Hugging Face's logo

llmfan46
/
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved

Image-Text-to-Text
Transformers
Safetensors
qwen3_5
heretic
uncensored
decensored
abliterated
mpoa
mtp
conversational
Model card Files
xet
 Community
2
llmfan46 commited on
Commit
a7c6845
·
verified ·
1 Parent(s): 0b2efbd

Upload folder using huggingface_hub

Browse files
Files changed (23) hide show
  1. .gitattributes +1 -0
  2. README.md +1256 -0
  3. chat_template.jinja +222 -0
  4. config.json +142 -0
  5. generation_config.json +13 -0
  6. model-00001-of-00012.safetensors +3 -0
  7. model-00002-of-00012.safetensors +3 -0
  8. model-00003-of-00012.safetensors +3 -0
  9. model-00004-of-00012.safetensors +3 -0
  10. model-00005-of-00012.safetensors +3 -0
  11. model-00006-of-00012.safetensors +3 -0
  12. model-00007-of-00012.safetensors +3 -0
  13. model-00008-of-00012.safetensors +3 -0
  14. model-00009-of-00012.safetensors +3 -0
  15. model-00010-of-00012.safetensors +3 -0
  16. model-00011-of-00012.safetensors +3 -0
  17. model-00012-of-00012.safetensors +3 -0
  18. model-auxiliary.safetensors +3 -0
  19. model.safetensors.index.json +0 -0
  20. preprocessor_config.json +21 -0
  21. tokenizer.json +3 -0
  22. tokenizer_config.json +32 -0
  23. video_preprocessor_config.json +21 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,1256 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ license: apache-2.0
4
+ license_link: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE
5
+ pipeline_tag: image-text-to-text
6
+ tags:
7
+ - heretic
8
+ - uncensored
9
+ - decensored
10
+ - abliterated
11
+ - mpoa
12
+ base_model:
13
+ - Qwen/Qwen3.6-27B
14
+ ---
15
+ <div style="background-color: #ff4444; color: white; padding: 20px; border-radius: 10px; text-align: center; margin: 20px 0;">
16
+ <h2 style="color: white; margin: 0 0 10px 0;">🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨</h2>
17
+ <p style="font-size: 18px; margin: 0 0 15px 0;">I can no longer upload new models unless I can cover the cost of additional storage.<br>I host <b>70+ free models</b> as an independent contributor and this work is unpaid.<br><b>Without your support, no more new models can be uploaded.</b></p>
18
+ <p style="font-size: 20px; margin: 0;">
19
+ <a href="https://patreon.com/LLMfan46" style="color: white; text-decoration: underline;">🎉 Patreon (Monthly)</a> &nbsp;|&nbsp;
20
+ <a href="https://ko-fi.com/llmfan46" style="color: white; text-decoration: underline;">☕ Ko-fi (One-time)</a>
21
+ </p>
22
+ <p style="font-size: 16px; margin: 10px 0 0 0;">Every contribution goes directly toward Hugging Face storage fees to keep models free for everyone.</p>
23
+ </div>
24
+
25
+ ---
26
+
27
+ This is the full model with the 15 MTPs all intacts.
28
+
29
+ ### **94% fewer refusals** (6/100 Uncensored vs 92/100 Original) while preserving model quality (0.0021 KL divergence).
30
+
31
+ ## ❤️ Support My Work
32
+ Creating these models takes significant time, work and compute. If you find them useful consider supporting me:
33
+
34
+ ![image/png](https://huggingface.co/llmfan46/Omega-Darker-Gaslight_The-Final-Forgotten-Fever-Dream-24B-ultra-uncensored-heretic-v1/resolve/main/waifu001.webp)
35
+
36
+ | Platform | Link | What you get |
37
+ |----------|------|--------------|
38
+ | 🎉 Patreon | [Monthly support](https://patreon.com/LLMfan46) | Priority model requests |
39
+ | ☕ Ko-fi | [One-time tip](https://ko-fi.com/llmfan46) | My eternal gratitude |
40
+
41
+ Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.
42
+
43
+ -----
44
+
45
+ # This is a decensored version of [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B), made using [Heretic](https://github.com/p-e-w/heretic) v1.2.0 with a variant of the [Magnitude-Preserving Orthogonal Ablation (MPOA)](https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration) method
46
+
47
+ ## Abliteration parameters
48
+
49
+ | Parameter | Value |
50
+ | :-------- | :---: |
51
+ | **direction_index** | 30.38 |
52
+ | **attn.out_proj.max_weight** | 1.58 |
53
+ | **attn.out_proj.max_weight_position** | 38.93 |
54
+ | **attn.out_proj.min_weight** | 1.51 |
55
+ | **attn.out_proj.min_weight_distance** | 32.78 |
56
+ | **mlp.down_proj.max_weight** | 1.80 |
57
+ | **mlp.down_proj.max_weight_position** | 41.28 |
58
+ | **mlp.down_proj.min_weight** | 0.54 |
59
+ | **mlp.down_proj.min_weight_distance** | 43.66 |
60
+ | **attn.o_proj.max_weight** | 1.99 |
61
+ | **attn.o_proj.max_weight_position** | 48.06 |
62
+ | **attn.o_proj.min_weight** | 1.75 |
63
+ | **attn.o_proj.min_weight_distance** | 39.00 |
64
+
65
+ ## Targeted components
66
+
67
+ * attn.o_proj
68
+ * attn.out_proj
69
+ * mlp.down_proj
70
+
71
+ ## Performance
72
+
73
+ | Metric | This model | Original model ([Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B)) |
74
+ | :----- | :--------: | :---------------------------: |
75
+ | **KL divergence** | <span style="color:darkgoldenrod">0.0021</span> | 0 *(by definition)* |
76
+ | **Refusals** | ✅ <span style="color:darkgreen">6/100</span> | ❌ <span style="color:blue">92/100</span> |
77
+
78
+ Lower refusals indicate fewer content restrictions, while lower KL divergence indicates more closeness to the original model's baseline. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections.
79
+
80
+ ## MMLU test results:
81
+
82
+ <span style="color:blue">Original:</span>
83
+
84
+ ============================================================
85
+
86
+ - Total questions: 7021
87
+
88
+ - Correct: 6084
89
+
90
+ - Accuracy: 0.8665 (86.65%)
91
+
92
+ - Parse failures: 0
93
+
94
+ ============================================================
95
+
96
+ Tested subject scores:
97
+ - professional_law: 0.7580 (595/785)
98
+ - moral_scenarios: 0.7715 (341/442)
99
+ - miscellaneous: 0.9399 (360/383)
100
+ - professional_psychology: 0.8987 (284/316)
101
+ - high_school_psychology: 0.9704 (262/270)
102
+ - high_school_macroeconomics: 0.9289 (183/197)
103
+ - elementary_mathematics: 0.8967 (165/184)
104
+ - moral_disputes: 0.8621 (150/174)
105
+ - prehistory: 0.8953 (154/172)
106
+ - philosophy: 0.8868 (141/159)
107
+ - high_school_biology: 0.9539 (145/152)
108
+ - professional_accounting: 0.8531 (122/143)
109
+ - clinical_knowledge: 0.9071 (127/140)
110
+ - high_school_microeconomics: 0.9706 (132/136)
111
+ - nutrition: 0.9111 (123/135)
112
+ - professional_medicine: 0.9478 (127/134)
113
+ - conceptual_physics: 0.9141 (117/128)
114
+ - high_school_mathematics: 0.6378 (81/127)
115
+ - human_aging: 0.8190 (95/116)
116
+ - security_studies: 0.9107 (102/112)
117
+ - high_school_statistics: 0.9009 (100/111)
118
+ - marketing: 0.9633 (105/109)
119
+ - high_school_world_history: 0.9623 (102/106)
120
+ - sociology: 0.9417 (97/103)
121
+ - high_school_government_and_politics: 0.9802 (99/101)
122
+ - high_school_geography: 0.9697 (96/99)
123
+ - high_school_chemistry: 0.8041 (78/97)
124
+ - high_school_us_history: 0.9895 (94/95)
125
+ - virology: 0.5506 (49/89)
126
+ - college_medicine: 0.8409 (74/88)
127
+ - world_religions: 0.9091 (80/88)
128
+ - high_school_physics: 0.8571 (72/84)
129
+ - electrical_engineering: 0.8642 (70/81)
130
+ - astronomy: 0.9747 (77/79)
131
+ - logical_fallacies: 0.9342 (71/76)
132
+ - high_school_european_history: 0.9452 (69/73)
133
+ - anatomy: 0.8169 (58/71)
134
+ - college_biology: 0.9375 (60/64)
135
+ - human_sexuality: 0.8906 (57/64)
136
+ - formal_logic: 0.7969 (51/64)
137
+ - public_relations: 0.7377 (45/61)
138
+ - international_law: 0.9000 (54/60)
139
+ - college_physics: 0.7719 (44/57)
140
+ - college_mathematics: 0.7636 (42/55)
141
+ - econometrics: 0.7963 (43/54)
142
+ - jurisprudence: 0.9057 (48/53)
143
+ - high_school_computer_science: 0.9615 (50/52)
144
+ - machine_learning: 0.8269 (43/52)
145
+ - medical_genetics: 0.9412 (48/51)
146
+ - global_facts: 0.6275 (32/51)
147
+ - management: 0.9000 (45/50)
148
+ - us_foreign_policy: 0.9600 (48/50)
149
+ - college_chemistry: 0.7021 (33/47)
150
+ - abstract_algebra: 0.7021 (33/47)
151
+ - business_ethics: 0.8261 (38/46)
152
+ - college_computer_science: 0.8222 (37/45)
153
+ - computer_security: 0.8372 (36/43)
154
+
155
+
156
+ <span style="color:darkgreen">Heretic:</span>
157
+
158
+ ============================================================
159
+
160
+ - Total questions: 7021
161
+
162
+ - Correct: 6015
163
+
164
+ - Accuracy: 0.8567 (85.67%)
165
+
166
+ - Parse failures: 0
167
+
168
+ ============================================================
169
+
170
+ Tested subject scores:
171
+ - professional_law: 0.7146 (561/785)
172
+ - moral_scenarios: 0.7466 (331/442)
173
+ - miscellaneous: 0.9347 (361/383)
174
+ - professional_psychology: 0.9019 (284/316)
175
+ - high_school_psychology: 0.9704 (262/270)
176
+ - high_school_macroeconomics: 0.9391 (185/197)
177
+ - elementary_mathematics: 0.8967 (164/184)
178
+ - moral_disputes: 0.8621 (149/174)
179
+ - prehistory: 0.8779 (151/172)
180
+ - philosophy: 0.8931 (141/159)
181
+ - high_school_biology: 0.9539 (144/152)
182
+ - professional_accounting: 0.8182 (118/143)
183
+ - clinical_knowledge: 0.9143 (128/140)
184
+ - high_school_microeconomics: 0.9706 (132/136)
185
+ - nutrition: 0.8815 (120/135)
186
+ - professional_medicine: 0.9478 (127/134)
187
+ - conceptual_physics: 0.9141 (117/128)
188
+ - high_school_mathematics: 0.6299 (82/127)
189
+ - human_aging: 0.8103 (94/116)
190
+ - security_studies: 0.8661 (97/112)
191
+ - high_school_statistics: 0.8829 (98/111)
192
+ - marketing: 0.9633 (105/109)
193
+ - high_school_world_history: 0.9528 (101/106)
194
+ - sociology: 0.9417 (97/103)
195
+ - high_school_government_and_politics: 0.9604 (97/101)
196
+ - high_school_geography: 0.9697 (96/99)
197
+ - high_school_chemistry: 0.8041 (77/97)
198
+ - high_school_us_history: 0.9474 (90/95)
199
+ - virology: 0.5281 (47/89)
200
+ - college_medicine: 0.8523 (75/88)
201
+ - world_religions: 0.9205 (81/88)
202
+ - high_school_physics: 0.8571 (73/84)
203
+ - electrical_engineering: 0.8395 (69/81)
204
+ - astronomy: 0.9747 (77/79)
205
+ - logical_fallacies: 0.9342 (71/76)
206
+ - high_school_european_history: 0.9452 (69/73)
207
+ - anatomy: 0.8169 (58/71)
208
+ - college_biology: 0.9531 (61/64)
209
+ - human_sexuality: 0.8750 (56/64)
210
+ - formal_logic: 0.7812 (50/64)
211
+ - public_relations: 0.7377 (45/61)
212
+ - international_law: 0.9167 (55/60)
213
+ - college_physics: 0.8070 (46/57)
214
+ - college_mathematics: 0.7455 (40/55)
215
+ - econometrics: 0.8333 (45/54)
216
+ - jurisprudence: 0.8868 (47/53)
217
+ - high_school_computer_science: 0.9615 (50/52)
218
+ - machine_learning: 0.7692 (40/52)
219
+ - medical_genetics: 0.9412 (48/51)
220
+ - global_facts: 0.6471 (33/51)
221
+ - management: 0.9000 (45/50)
222
+ - us_foreign_policy: 0.9800 (49/50)
223
+ - college_chemistry: 0.6596 (31/47)
224
+ - abstract_algebra: 0.7021 (33/47)
225
+ - business_ethics: 0.7826 (37/46)
226
+ - college_computer_science: 0.8444 (38/45)
227
+ - computer_security: 0.8605 (37/43)
228
+
229
+ MMLU - Massive Multitask Language Understanding, multiple-choice questions across 57 subjects (math, history, law, medicine, etc.).
230
+
231
+ ## GGUF Version (non-MTP)
232
+
233
+ GGUF quantizations available here [llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF](https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF).
234
+
235
+ ## NVFP4 GGUF Version (non-MTP)
236
+
237
+ NVFP4 GGUF quantizations available here [llmfan46/Qwen3.6-27B-uncensored-heretic-v2-NVFP4-GGUF](https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-NVFP4-GGUF).
238
+
239
+ ## GPTQ-Int4 Version (non-MTP)
240
+
241
+ GPTQ 4-bit quantization available here [llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GPTQ-Int4](https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GPTQ-Int4).
242
+
243
+ ## GPTQ-Int8 Version (non-MTP)
244
+
245
+ GPTQ 8-bit quantization available here [llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GPTQ-Int8](https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GPTQ-Int8).
246
+
247
+ ## FP8-W8A16 Version (non-MTP)
248
+
249
+ FP8-W8A16 / float8_e4m3fn weight-only quantization available here [llmfan46/Qwen3.6-27B-uncensored-heretic-v2-FP8-W8A16](https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-FP8-W8A16).
250
+
251
+ -----
252
+
253
+
254
+ # Qwen3.6-27B
255
+
256
+ <img width="400px" src="https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.6/logo.png">
257
+
258
+ [![Qwen Chat](https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5)](https://chat.qwen.ai)
259
+
260
+ > [!Note]
261
+ > This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.
262
+ >
263
+ > These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.
264
+
265
+ Following the February release of the Qwen3.5 series, we're pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience.
266
+
267
+ ## Qwen3.6 Highlights
268
+
269
+ This release delivers substantial upgrades, particularly in
270
+
271
+ - **Agentic Coding:** the model now handles frontend workflows and repository-level reasoning with greater fluency and precision.
272
+ - **Thinking Preservation:** we've introduced a new option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.
273
+
274
+ ![Benchmark Results](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.6/Figures/qwen3.6_27b_score.png)
275
+
276
+ For more details, please refer to our blog post [Qwen3.6-27B](https://qwen.ai/blog?id=qwen3.6-27b).
277
+
278
+ ## Model Overview
279
+
280
+ - Type: Causal Language Model with Vision Encoder
281
+ - Training Stage: Pre-training & Post-training
282
+ - Language Model
283
+ - Number of Parameters: 27B
284
+ - Hidden Dimension: 5120
285
+ - Token Embedding: 248320 (Padded)
286
+ - Number of Layers: 64
287
+ - Hidden Layout: 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
288
+ - Gated DeltaNet:
289
+ - Number of Linear Attention Heads: 48 for V and 16 for QK
290
+ - Head Dimension: 128
291
+ - Gated Attention:
292
+ - Number of Attention Heads: 24 for Q and 4 for KV
293
+ - Head Dimension: 256
294
+ - Rotary Position Embedding Dimension: 64
295
+ - Feed Forward Network:
296
+ - Intermediate Dimension: 17408
297
+ - LM Output: 248320 (Padded)
298
+ - MTP: trained with multi-steps
299
+ - Context Length: 262,144 natively and extensible up to 1,010,000 tokens.
300
+
301
+
302
+ ## Benchmark Results
303
+
304
+ ### Language
305
+
306
+ <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;max-width:1000px;margin:0 auto;padding:16px 0">
307
+ <table style="width:100%;border-collapse:collapse;font-size:13px">
308
+ <thead><tr>
309
+ <th style="padding:10px 7px;text-align:left;font-weight:600;border-bottom:2px solid #7c3aed;color:#7c3aed"></th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-27B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-397B-A17B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Gemma4-31B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Claude 4.5 Opus</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-35B-A3B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-27B</th></tr></thead>
310
+ <tbody>
311
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Coding Agent</td></tr>
312
+ <tr>
313
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Verified</td>
314
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.0</td>
315
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.2</td>
316
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.0</td>
317
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.9</td>
318
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.4</td>
319
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.2</td>
320
+ </tr>
321
+ <tr>
322
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Pro</td>
323
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.2</td>
324
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.9</td>
325
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">35.7</td>
326
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.1</td>
327
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49.5</td>
328
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.5</td>
329
+ </tr>
330
+ <tr>
331
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Multilingual</td>
332
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>
333
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>
334
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.7</td>
335
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.5</td>
336
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.2</td>
337
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">71.3</td>
338
+ </tr>
339
+ <tr>
340
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Terminal-Bench 2.0</td>
341
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.6</td>
342
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.5</td>
343
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.9</td>
344
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.3</td>
345
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.5</td>
346
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.3</td>
347
+ </tr>
348
+ <tr>
349
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SkillsBench <sub><small>Avg5</small></sub></td>
350
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">27.2</td>
351
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">30.0</td>
352
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">23.6</td>
353
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">45.3</td>
354
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.7</td>
355
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.2</td>
356
+ </tr>
357
+ <tr>
358
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">QwenWebBench</td>
359
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1068</td>
360
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1186</td>
361
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1197</td>
362
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1536</td>
363
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1397</td>
364
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1487</td>
365
+ </tr>
366
+ <tr>
367
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">NL2Repo</td>
368
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">27.3</td>
369
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">32.2</td>
370
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">15.5</td>
371
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">43.2</td>
372
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">29.4</td>
373
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">36.2</td>
374
+ </tr>
375
+ <tr>
376
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Claw-Eval <sub><small>Avg</small></sub></td>
377
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.3</td>
378
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.7</td>
379
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.5</td>
380
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.6</td>
381
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.7</td>
382
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.4</td>
383
+ </tr>
384
+ <tr>
385
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Claw-Eval <sub><small>Pass^3</small></sub></td>
386
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.2</td>
387
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.1</td>
388
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">25.0</td>
389
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.6</td>
390
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.0</td>
391
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">60.6</td>
392
+ </tr>
393
+ <tr>
394
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">QwenClawBench</td>
395
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.2</td>
396
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.8</td>
397
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.7</td>
398
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.3</td>
399
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.6</td>
400
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.4</td>
401
+ </tr>
402
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Knowledge</td></tr>
403
+ <tr>
404
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Pro</td>
405
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.1</td>
406
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.8</td>
407
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.2</td>
408
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.5</td>
409
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.2</td>
410
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.2</td>
411
+ </tr>
412
+ <tr>
413
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Redux</td>
414
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.2</td>
415
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.9</td>
416
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.7</td>
417
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">95.6</td>
418
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.3</td>
419
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.5</td>
420
+ </tr>
421
+ <tr>
422
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SuperGPQA</td>
423
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.6</td>
424
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.4</td>
425
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.7</td>
426
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.6</td>
427
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.7</td>
428
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">66.0</td>
429
+ </tr>
430
+ <tr>
431
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">C-Eval</td>
432
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.5</td>
433
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.0</td>
434
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.6</td>
435
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.2</td>
436
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.0</td>
437
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">91.4</td>
438
+ </tr>
439
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">STEM & Reasoning</td></tr>
440
+ <tr>
441
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">GPQA Diamond</td>
442
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.5</td>
443
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">88.4</td>
444
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>
445
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.0</td>
446
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.0</td>
447
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.8</td>
448
+ </tr>
449
+ <tr>
450
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HLE</td>
451
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">24.3</td>
452
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.7</td>
453
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">19.5</td>
454
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">30.8</td>
455
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">21.4</td>
456
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">24.0</td>
457
+ </tr>
458
+ <tr>
459
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">LiveCodeBench v6</td>
460
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>
461
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.6</td>
462
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.0</td>
463
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.8</td>
464
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.4</td>
465
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.9</td>
466
+ </tr>
467
+ <tr>
468
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Feb 25</td>
469
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.0</td>
470
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.8</td>
471
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">88.7</td>
472
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.9</td>
473
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.7</td>
474
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.8</td>
475
+ </tr>
476
+ <tr>
477
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Nov 25</td>
478
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.8</td>
479
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.7</td>
480
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.5</td>
481
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.3</td>
482
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.1</td>
483
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.7</td>
484
+ </tr>
485
+ <tr>
486
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Feb 26</td>
487
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>
488
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.9</td>
489
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.2</td>
490
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.3</td>
491
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.6</td>
492
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>
493
+ </tr>
494
+ <tr>
495
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">IMOAnswerBench</td>
496
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.9</td>
497
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.9</td>
498
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.5</td>
499
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.0</td>
500
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.9</td>
501
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.8</td>
502
+ </tr>
503
+ <tr>
504
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">AIME26</td>
505
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.6</td>
506
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.3</td>
507
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.2</td>
508
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">95.1</td>
509
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.7</td>
510
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.1</td>
511
+ </tr>
512
+ </tbody>
513
+ </table>
514
+
515
+ <p style="margin-top:12px;font-size:10px;opacity:0.7">
516
+ * SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top_p=0.95, 200K context window. We correct some problematic tasks in the public set of SWE-bench Pro and evaluate all baselines on the refined benchmark.<br/>
517
+ * Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs.<br/>
518
+ * SkillsBench: Evaluated via OpenCode on 78 tasks (self-contained subset, excluding API-dependent tasks); avg of 5 runs.<br/>
519
+ * NL2Repo: Others are evaluated via Claude Code (temp=1.0, top_p=0.95, max_turns=900).<br/>
520
+ * QwenClawBench: A real-user-distribution Claw agent benchmark; temp=0.6, 256K ctx.<br/>
521
+ * QwenWebBench: An internal front-end code generation benchmark; bilingual (EN/CN), 7 categories (Web Design, Web Apps, Games, SVG, Data Visualization, Animation, and 3D); auto-render + multimodal judge (code/visual correctness); BT/Elo rating system.<br/>
522
+ * AIME 26: We use the full AIME 2026 (I & II), where the scores may differ from Qwen 3.5 notes.
523
+ </p>
524
+
525
+ </div>
526
+
527
+
528
+ ### Vision Language
529
+
530
+ <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;max-width:1000px;margin:0 auto;padding:16px 0">
531
+ <table style="width:100%;border-collapse:collapse;font-size:13px">
532
+ <thead><tr>
533
+ <th style="padding:10px 7px;text-align:left;font-weight:600;border-bottom:2px solid #7c3aed;color:#7c3aed"></th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-27B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-397B-A17B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Gemma4-31B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Claude 4.5 Opus</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-35B-A3B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-27B</th></tr></thead>
534
+ <tbody>
535
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">STEM & Puzzle</td></tr>
536
+ <tr>
537
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMU</td>
538
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.3</td>
539
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.0</td>
540
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.4</td>
541
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>
542
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.7</td>
543
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.9</td>
544
+ </tr>
545
+ <tr>
546
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMU-Pro</td>
547
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.0</td>
548
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.0</td>
549
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.9</td>
550
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.6</td>
551
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.3</td>
552
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.8</td>
553
+ </tr>
554
+ <tr>
555
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MathVista <sub><small>mini</small></sub></td>
556
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.8</td>
557
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
558
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.3</td>
559
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
560
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.4</td>
561
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.4</td>
562
+ </tr>
563
+ <tr>
564
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">DynaMath</td>
565
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.7</td>
566
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.3</td>
567
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.5</td>
568
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.7</td>
569
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.8</td>
570
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.6</td>
571
+ </tr>
572
+ <tr>
573
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VlmsAreBlind</td>
574
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.9</td>
575
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
576
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.2</td>
577
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
578
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.6</td>
579
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.0</td>
580
+ </tr>
581
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">General VQA</td></tr>
582
+ <tr>
583
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RealWorldQA</td>
584
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.7</td>
585
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.9</td>
586
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.3</td>
587
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.0</td>
588
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.3</td>
589
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.1</td>
590
+ </tr>
591
+ <tr>
592
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMStar</td>
593
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.0</td>
594
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.8</td>
595
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.3</td>
596
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.2</td>
597
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>
598
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.4</td>
599
+ </tr>
600
+ <tr>
601
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMBench<sub><small>EN-DEV-v1.1</small></sub></td>
602
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.6</td>
603
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
604
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.9</td>
605
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
606
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.8</td>
607
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.3</td>
608
+ </tr>
609
+ <tr>
610
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SimpleVQA</td>
611
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">56.0</td>
612
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.1</td>
613
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.9</td>
614
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.7</td>
615
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.9</td>
616
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">56.1</td>
617
+ </tr>
618
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Document Understanding</td></tr>
619
+ <tr>
620
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CharXiv <sub><small>RQ</small></sub></td>
621
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.5</td>
622
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.8</td>
623
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.9</td>
624
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.5</td>
625
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.0</td>
626
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.4</td>
627
+ </tr>
628
+ <tr>
629
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CC-OCR</td>
630
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.0</td>
631
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.0</td>
632
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.7</td>
633
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.9</td>
634
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.9</td>
635
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.2</td>
636
+ </tr>
637
+ <tr>
638
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">OCRBench</td>
639
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.4</td>
640
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
641
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.1</td>
642
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
643
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.0</td>
644
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.4</td>
645
+ </tr>
646
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Spatial Intelligence</td></tr>
647
+ <tr>
648
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ERQA</td>
649
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">60.5</td>
650
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.5</td>
651
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.5</td>
652
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.8</td>
653
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.8</td>
654
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.5</td>
655
+ </tr>
656
+ <tr>
657
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CountBench</td>
658
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.8</td>
659
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.2</td>
660
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.1</td>
661
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.6</td>
662
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.1</td>
663
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.8</td>
664
+ </tr>
665
+ <tr>
666
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RefCOCO <sub><small>avg</small></sub></td>
667
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.9</td>
668
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.3</td>
669
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
670
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
671
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.0</td>
672
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.5</td>
673
+ </tr>
674
+ <tr>
675
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">EmbSpatialBench</td>
676
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.5</td>
677
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
678
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
679
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
680
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>
681
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.6</td>
682
+ </tr>
683
+ <tr>
684
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RefSpatialBench</td>
685
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.7</td>
686
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
687
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">4.7</td>
688
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
689
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.3</td>
690
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.0</td>
691
+ </tr>
692
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Video Understanding</td></tr>
693
+ <tr>
694
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VideoMME<sub><small>(w sub.)</sub></small></td>
695
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.0</td>
696
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.5</td>
697
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
698
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.7</td>
699
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.6</td>
700
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.7</td>
701
+ </tr>
702
+ <tr>
703
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VideoMMMU</td>
704
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.3</td>
705
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.7</td>
706
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.6</td>
707
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.4</td>
708
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.7</td>
709
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.4</td>
710
+ </tr>
711
+ <tr>
712
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MLVU</td>
713
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.9</td>
714
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.7</td>
715
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
716
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.7</td>
717
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.2</td>
718
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.6</td>
719
+ </tr>
720
+ <tr>
721
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MVBench</td>
722
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.6</td>
723
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.6</td>
724
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
725
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.2</td>
726
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.6</td>
727
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.5</td>
728
+ </tr>
729
+ <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Visual Agent</td></tr>
730
+ <tr>
731
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">V*</td>
732
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.7</td>
733
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">95.8</td>
734
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
735
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.0</td>
736
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.1</td>
737
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.7</td>
738
+ </tr>
739
+ <tr>
740
+ <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">AndroidWorld</td>
741
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.2</td>
742
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
743
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
744
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
745
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
746
+ <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.3</td>
747
+ </tr>
748
+ </tbody>
749
+ </table>
750
+
751
+ <p style="margin-top:12px;font-size:10px;opacity:0.7">
752
+ * Empty cells (--) indicate scores not yet available or not applicable.
753
+ </p>
754
+
755
+ </div>
756
+
757
+
758
+ ## Quickstart
759
+
760
+ For streamlined integration, we recommend using Qwen3.6 via APIs. Below is a guide to use Qwen3.6 via OpenAI-compatible API.
761
+
762
+ ### Serving Qwen3.6
763
+
764
+ Qwen3.6 can be served via APIs with popular inference frameworks.
765
+ In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.6 models.
766
+
767
+ > [!Important]
768
+ > Inference efficiency and throughput vary significantly across frameworks.
769
+ > We recommend using the latest framework versions to ensure optimal performance and compatibility.
770
+ > For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.
771
+
772
+ > [!Important]
773
+ > The model has a default context length of 262,144 tokens.
774
+ > If you encounter out-of-memory (OOM) errors, consider reducing the context window.
775
+ > However, because Qwen3.6 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.
776
+
777
+ #### SGLang
778
+
779
+ [SGLang](https://github.com/sgl-project/sglang) is a fast serving framework for large language models and vision language models.
780
+ `sglang>=0.5.10` is recommended for Qwen3.6, which can be installed using the following command in a fresh environment:
781
+ ```shell
782
+ uv pip install sglang[all]
783
+ ```
784
+ See [its documentation](https://docs.sglang.ai/get_started/install.html) for more details.
785
+
786
+ The following will create API endpoints at `http://localhost:8000/v1`:
787
+
788
+ - **Standard Version**: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.
789
+
790
+ ```shell
791
+ python -m sglang.launch_server --model-path Qwen/Qwen3.6-27B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3
792
+ ```
793
+
794
+ - **Tool Use**: To support tool use, you can use the following command.
795
+
796
+ ```shell
797
+ python -m sglang.launch_server --model-path Qwen/Qwen3.6-27B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder
798
+ ```
799
+
800
+ - **Multi-Token Prediction (MTP)**: The following command is recommended for MTP:
801
+
802
+ ```shell
803
+ python -m sglang.launch_server --model-path Qwen/Qwen3.6-27B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
804
+ ```
805
+
806
+ For detailed deployment guide, see the [SGLang Qwen3.5 Cookbook](https://lmsysorg.mintlify.app/cookbook/llm/Qwen/Qwen3.5).
807
+
808
+ #### vLLM
809
+
810
+ [vLLM](https://github.com/vllm-project/vllm) is a high-throughput and memory-efficient inference and serving engine for LLMs.
811
+ `vllm>=0.19.0` is recommended for Qwen3.6, which can be installed using the following command in a fresh environment:
812
+ ```shell
813
+ uv pip install vllm --torch-backend=auto
814
+ ```
815
+ See [its documentation](https://docs.vllm.ai/en/stable/getting_started/installation/index.html) for more details.
816
+
817
+
818
+ The following will create API endpoints at `http://localhost:8000/v1`:
819
+
820
+ - **Standard Version**: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.
821
+
822
+ ```shell
823
+ vllm serve Qwen/Qwen3.6-27B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3
824
+ ```
825
+
826
+ - **Tool Call**: To support tool use, you can use the following command.
827
+
828
+ ```shell
829
+ vllm serve Qwen/Qwen3.6-27B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
830
+ ```
831
+
832
+ - **Multi-Token Prediction (MTP)**: The following command is recommended for MTP:
833
+
834
+ ```shell
835
+ vllm serve Qwen/Qwen3.6-27B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
836
+ ```
837
+
838
+ - **Text-Only**: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:
839
+
840
+ ```shell
841
+ vllm serve Qwen/Qwen3.6-27B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only
842
+ ```
843
+
844
+ For detailed deployment guide, see the [vLLM Qwen3.5 Recipe](https://docs.vllm.ai/projects/recipes/en/latest/Qwen/Qwen3.5.html).
845
+
846
+ #### KTransformers
847
+
848
+ [KTransformers](https://github.com/kvcache-ai/ktransformers) is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing.
849
+ For running Qwen3.6 with KTransformers, see the [KTransformers Deployment Guide](https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/Qwen3.5.md).
850
+
851
+ #### Hugging Face Transformers
852
+
853
+ Hugging Face Transformers contains a _lightweight_ server which can be used for quick testing and moderate load deployment.
854
+ The latest `transformers` is required for Qwen3.6:
855
+ ```shell
856
+ pip install "transformers[serving]"
857
+ ```
858
+ See [its documentation](https://huggingface.co/docs/transformers/main/serving) for more details. Please also make sure torchvision and pillow are installed.
859
+
860
+ Then, run `transformers serve` to launch a server with API endpoints at `http://localhost:8000/v1`; it will place the model on accelerators if available:
861
+ ```shell
862
+ transformers serve Qwen/Qwen3.6-27B --port 8000 --continuous-batching
863
+ ```
864
+
865
+ ### Using Qwen3.6 via the Chat Completions API
866
+
867
+ The chat completions API is accessible via standard HTTP requests or OpenAI SDKs.
868
+ Here, we show examples using the OpenAI Python SDK.
869
+
870
+ Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:
871
+ ```shell
872
+ pip install -U openai
873
+
874
+ # Set the following accordingly
875
+ export OPENAI_BASE_URL="http://localhost:8000/v1"
876
+ export OPENAI_API_KEY="EMPTY"
877
+ ```
878
+
879
+ > [!Tip]
880
+ > We recommend using the following set of sampling parameters for generation
881
+ > - Thinking mode for general tasks: `temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0`
882
+ > - Thinking mode for precise coding tasks (e.g. WebDev): `temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0`
883
+ > - Instruct (or non-thinking) mode: `temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0`
884
+ >
885
+ > Please note that the support for sampling parameters varies according to inference frameworks.
886
+
887
+ > [!Important]
888
+ > Qwen3.6 models operate in thinking mode by default, generating thinking content signified by `<think>\n...</think>\n\n` before producing the final responses.
889
+ > To disable thinking content and obtain direct response, refer to the examples [here](#instruct-or-non-thinking-mode).
890
+
891
+
892
+ #### Text-Only Input
893
+
894
+ ```python
895
+ from openai import OpenAI
896
+ # Configured by environment variables
897
+ client = OpenAI()
898
+
899
+ messages = [
900
+ {"role": "user", "content": "Type \"I love Qwen3.6\" backwards"},
901
+ ]
902
+
903
+ chat_response = client.chat.completions.create(
904
+ model="Qwen/Qwen3.6-27B",
905
+ messages=messages,
906
+ max_tokens=81920,
907
+ temperature=1.0,
908
+ top_p=0.95,
909
+ presence_penalty=0.0,
910
+ extra_body={
911
+ "top_k": 20,
912
+ },
913
+ )
914
+ print("Chat response:", chat_response)
915
+ ```
916
+
917
+
918
+ #### Image Input
919
+
920
+ ```python
921
+ from openai import OpenAI
922
+ # Configured by environment variables
923
+ client = OpenAI()
924
+
925
+ messages = [
926
+ {
927
+ "role": "user",
928
+ "content": [
929
+ {
930
+ "type": "image_url",
931
+ "image_url": {
932
+ "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"
933
+ }
934
+ },
935
+ {
936
+ "type": "text",
937
+ "text": "The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"
938
+ }
939
+ ]
940
+ }
941
+ ]
942
+
943
+ response = client.chat.completions.create(
944
+ model="Qwen/Qwen3.6-27B",
945
+ messages=messages,
946
+ max_tokens=81920,
947
+ temperature=1.0,
948
+ top_p=0.95,
949
+ presence_penalty=0.0,
950
+ extra_body={
951
+ "top_k": 20,
952
+ },
953
+ )
954
+ print("Chat response:", chat_response)
955
+ ```
956
+
957
+ #### Video Input
958
+
959
+ ```python
960
+ from openai import OpenAI
961
+ # Configured by environment variables
962
+ client = OpenAI()
963
+
964
+ messages = [
965
+ {
966
+ "role": "user",
967
+ "content": [
968
+ {
969
+ "type": "video_url",
970
+ "video_url": {
971
+ "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
972
+ }
973
+ },
974
+ {
975
+ "type": "text",
976
+ "text": "How many porcelain jars were discovered in the niches located in the primary chamber of the tomb?"
977
+ }
978
+ ]
979
+ }
980
+ ]
981
+
982
+ # When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
983
+ # video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
984
+ # This feature is currently supported only in vLLM.
985
+ #
986
+ # By default, `fps=2` and `do_sample_frames=True`.
987
+ # With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
988
+ response = client.chat.completions.create(
989
+ model="Qwen/Qwen3.6-27B",
990
+ messages=messages,
991
+ max_tokens=81920,
992
+ temperature=1.0,
993
+ top_p=0.95,
994
+ presence_penalty=0.0,
995
+ extra_body={
996
+ "top_k": 20,
997
+ "mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
998
+ },
999
+ )
1000
+
1001
+ print("Chat response:", chat_response)
1002
+ ```
1003
+
1004
+
1005
+ #### Instruct (or Non-Thinking) Mode
1006
+
1007
+ > [!Important]
1008
+ > Qwen3.6 does not officially support the soft switch of Qwen3, i.e., `/think` and `/nothink`.
1009
+
1010
+ Qwen3.6 will think by default before response.
1011
+ You can obtain direct response from the model without thinking by configuring the API parameters.
1012
+ For example,
1013
+ ```python
1014
+ from openai import OpenAI
1015
+ # Configured by environment variables
1016
+ client = OpenAI()
1017
+
1018
+ messages = [
1019
+ {
1020
+ "role": "user",
1021
+ "content": [
1022
+ {
1023
+ "type": "image_url",
1024
+ "image_url": {
1025
+ "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.6/demo/RealWorld/RealWorld-04.png"
1026
+ }
1027
+ },
1028
+ {
1029
+ "type": "text",
1030
+ "text": "Where is this?"
1031
+ }
1032
+ ]
1033
+ }
1034
+ ]
1035
+
1036
+ chat_response = client.chat.completions.create(
1037
+ model="Qwen/Qwen3.6-27B",
1038
+ messages=messages,
1039
+ max_tokens=32768,
1040
+ temperature=0.7,
1041
+ top_p=0.8,
1042
+ presence_penalty=1.5,
1043
+ extra_body={
1044
+ "top_k": 20,
1045
+ "chat_template_kwargs": {"enable_thinking": False},
1046
+ },
1047
+ )
1048
+ print("Chat response:", chat_response)
1049
+ ```
1050
+
1051
+ > [!Note]
1052
+ > If you are using APIs from Alibaba Cloud Model Studio, in addition to changing `model`, please use `"enable_thinking": False` instead of `"chat_template_kwargs": {"enable_thinking": False}`.
1053
+
1054
+ #### Preserve Thinking
1055
+
1056
+ By default, only the thinking blocks generated in handling the latest user message is retained, resulting in a pattern commonly as interleaved thinking.
1057
+ Qwen3.6 has been additionally trained to preserve and leverage thinking traces from historical messages.
1058
+ You can enable this behavior by setting the `preserve_thinking` option:
1059
+ ```python
1060
+ from openai import OpenAI
1061
+ # Configured by environment variables
1062
+ client = OpenAI()
1063
+
1064
+ messages = [...]
1065
+
1066
+ chat_response = client.chat.completions.create(
1067
+ model="Qwen/Qwen3.6-27B",
1068
+ messages=messages,
1069
+ max_tokens=32768,
1070
+ temperature=0.6,
1071
+ top_p=0.95,
1072
+ presence_penalty=0.0,
1073
+ extra_body={
1074
+ "top_k": 20,
1075
+ "chat_template_kwargs": {"preserve_thinking": True},
1076
+ },
1077
+ )
1078
+ print("Chat response:", chat_response)
1079
+ ```
1080
+
1081
+ > [!Note]
1082
+ > If you are using APIs from Alibaba Cloud Model Studio, in addition to changing `model`, please use `"preserve_thinking": True` instead of `"chat_template_kwargs": {"preserve_thinking": False}`.
1083
+
1084
+
1085
+ This capability is particularly beneficial for agent scenarios, where maintaining full reasoning context can enhance decision consistency and, in many cases, reduce overall token consumption by minimizing redundant reasoning. Additionally, it can improve KV cache utilization, optimizing inference efficiency in both thinking and non-thinking modes.
1086
+
1087
+
1088
+ ## Agentic Usage
1089
+
1090
+ Qwen3.6 excels in tool calling capabilities.
1091
+
1092
+ ### Qwen-Agent
1093
+
1094
+ We recommend using [Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) to quickly build Agent applications with Qwen3.6.
1095
+
1096
+ To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.
1097
+ ```python
1098
+ import os
1099
+ from qwen_agent.agents import Assistant
1100
+
1101
+ # Define LLM
1102
+ # Using Alibaba Cloud Model Studio
1103
+ llm_cfg = {
1104
+ # Use the OpenAI-compatible model service provided by DashScope:
1105
+ 'model': 'qwen3.6-27b',
1106
+ 'model_type': 'qwenvl_oai',
1107
+ 'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
1108
+ 'api_key': os.getenv('DASHSCOPE_API_KEY'),
1109
+
1110
+ 'generate_cfg': {
1111
+ 'use_raw_api': True,
1112
+ # When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way
1113
+ 'extra_body': {
1114
+ 'enable_thinking': True,
1115
+ 'preserve_thinking': True,
1116
+ },
1117
+ },
1118
+ }
1119
+
1120
+ # Using OpenAI-compatible API endpoint.
1121
+ # functionality of the deployment frameworks and let Qwen-Agent automate the related operations.
1122
+ #
1123
+ # llm_cfg = {
1124
+ # # Use your own model service compatible with OpenAI API by vLLM/SGLang:
1125
+ # 'model': 'Qwen/Qwen3.6-27B',
1126
+ # 'model_type': 'qwenvl_oai',
1127
+ # 'model_server': 'http://localhost:8000/v1', # api_base
1128
+ # 'api_key': 'EMPTY',
1129
+ #
1130
+ # 'generate_cfg': {
1131
+ # 'use_raw_api': True,
1132
+ # # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way
1133
+ # 'extra_body': {
1134
+ # 'chat_template_kwargs': {'enable_thinking': True, 'preserve_thinking': True}
1135
+ # },
1136
+ # },
1137
+ # }
1138
+
1139
+ # Define Tools
1140
+ tools = [
1141
+ {'mcpServers': { # You can specify the MCP configuration file
1142
+ "filesystem": {
1143
+ "command": "npx",
1144
+ "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
1145
+ }
1146
+ }
1147
+ }
1148
+ ]
1149
+
1150
+ # Define Agent
1151
+ bot = Assistant(llm=llm_cfg, function_list=tools)
1152
+
1153
+ # Streaming generation
1154
+ messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
1155
+ for responses in bot.run(messages=messages):
1156
+ pass
1157
+ print(responses)
1158
+
1159
+ # Streaming generation
1160
+ messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
1161
+ for responses in bot.run(messages=messages):
1162
+ pass
1163
+ print(responses)
1164
+ ```
1165
+
1166
+ ### Qwen Code
1167
+
1168
+
1169
+ [Qwen Code](https://github.com/QwenLM/qwen-code) is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.
1170
+
1171
+ For more information, please refer to [Qwen Code](https://qwenlm.github.io/qwen-code-docs/).
1172
+
1173
+ ## Processing Ultra-Long Texts
1174
+
1175
+ Qwen3.6 natively supports context lengths of up to 262,144 tokens.
1176
+ For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.
1177
+
1178
+ YaRN is currently supported by several inference frameworks, e.g., `transformers`, `vllm`, `ktransformers` and `sglang`.
1179
+ In general, there are two approaches to enabling YaRN for supported frameworks:
1180
+
1181
+ - Modifying the model configuration file:
1182
+ In the `config.json` file, change the `rope_parameters` fields in `text_config` to:
1183
+ ```json
1184
+ {
1185
+ "mrope_interleaved": true,
1186
+ "mrope_section": [
1187
+ 11,
1188
+ 11,
1189
+ 10
1190
+ ],
1191
+ "rope_type": "yarn",
1192
+ "rope_theta": 10000000,
1193
+ "partial_rotary_factor": 0.25,
1194
+ "factor": 4.0,
1195
+ "original_max_position_embeddings": 262144,
1196
+ }
1197
+ ```
1198
+
1199
+ - Passing command line arguments:
1200
+
1201
+ For `vllm`, you can use
1202
+ ```shell
1203
+ VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000
1204
+ ```
1205
+
1206
+ For `sglang` and `ktransformers`, you can use
1207
+ ```shell
1208
+ SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000
1209
+ ```
1210
+
1211
+ > [!NOTE]
1212
+ > All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, **potentially impacting performance on shorter texts.**
1213
+ > We advise modifying the `rope_parameters` configuration only when processing long contexts is required.
1214
+ > It is also recommended to modify the `factor` as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set `factor` as 2.0.
1215
+
1216
+ ## Best Practices
1217
+
1218
+ To achieve optimal performance, we recommend the following settings:
1219
+
1220
+ 1. **Sampling Parameters**:
1221
+ - We suggest using the following sets of sampling parameters depending on the mode and task type:
1222
+ - **Thinking mode for general tasks**:
1223
+ `temperature=1.0`, `top_p=0.95`, `top_k=20`, `min_p=0.0`, `presence_penalty=0.0`, `repetition_penalty=1.0`
1224
+ - **Thinking mode for precise coding tasks (e.g., WebDev)**:
1225
+ `temperature=0.6`, `top_p=0.95`, `top_k=20`, `min_p=0.0`, `presence_penalty=0.0`, `repetition_penalty=1.0`
1226
+ - **Instruct (or non-thinking) mode**:
1227
+ `temperature=0.7`, `top_p=0.80`, `top_k=20`, `min_p=0.0`, `presence_penalty=1.5`, `repetition_penalty=1.0`
1228
+ - For supported frameworks, you can adjust the `presence_penalty` parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.
1229
+
1230
+ 2. **Adequate Output Length**: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.
1231
+
1232
+ 3. **Standardize Output Format**: We recommend using prompts to standardize model outputs when benchmarking.
1233
+ - **Math Problems**: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
1234
+ - **Multiple-Choice Questions**: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the `answer` field with only the choice letter, e.g., `"answer": "C"`."
1235
+
1236
+ 4. **Long Video Understanding**: To optimize inference efficiency for plain text and images, the `size` parameter in the released `video_preprocessor_config.json` is conservatively configured. It is recommended to set the `longest_edge` parameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example,
1237
+ ```json
1238
+ {"longest_edge": 469762048, "shortest_edge": 4096}
1239
+ ```
1240
+
1241
+ Alternatively, override the default values via engine startup parameters. For implementation details, refer to: [vLLM](https://github.com/vllm-project/vllm/pull/34330) / [SGLang](https://github.com/sgl-project/sglang/pull/18467).
1242
+
1243
+
1244
+ ### Citation
1245
+
1246
+ If you find our work helpful, feel free to give us a cite.
1247
+
1248
+ ```bibtex
1249
+ @misc{qwen3.6-27b,
1250
+ title = {{Qwen3.6-27B}: Flagship-Level Coding in a {27B} Dense Model},
1251
+ author = {{Qwen Team}},
1252
+ month = {April},
1253
+ year = {2026},
1254
+ url = {https://qwen.ai/blog?id=qwen3.6-27b}
1255
+ }
1256
+ ```
chat_template.jinja ADDED
@@ -0,0 +1,222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- set image_count = namespace(value=0) %}
2
+ {%- set video_count = namespace(value=0) %}
3
+ {%- macro render_content(content, do_vision_count, is_system_content=false) %}
4
+ {%- if content is string %}
5
+ {{- content }}
6
+ {%- elif content is iterable and content is not mapping %}
7
+ {%- for item in content %}
8
+ {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
9
+ {%- if is_system_content %}
10
+ {{- raise_exception('System message cannot contain images.') }}
11
+ {%- endif %}
12
+ {%- if do_vision_count %}
13
+ {%- set image_count.value = image_count.value + 1 %}
14
+ {%- endif %}
15
+ {%- if add_vision_id is defined and add_vision_id %}
16
+ {{- 'Picture ' ~ image_count.value ~ ': ' }}
17
+ {%- endif %}
18
+ {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
19
+ {%- elif 'video' in item or item.type == 'video' %}
20
+ {%- if is_system_content %}
21
+ {{- raise_exception('System message cannot contain videos.') }}
22
+ {%- endif %}
23
+ {%- if do_vision_count %}
24
+ {%- set video_count.value = video_count.value + 1 %}
25
+ {%- endif %}
26
+ {%- if add_vision_id is defined and add_vision_id %}
27
+ {{- 'Video ' ~ video_count.value ~ ': ' }}
28
+ {%- endif %}
29
+ {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
30
+ {%- elif 'text' in item %}
31
+ {{- item.text }}
32
+ {%- else %}
33
+ {{- raise_exception('Unexpected item type in content.') }}
34
+ {%- endif %}
35
+ {%- endfor %}
36
+ {%- elif content is none or content is undefined %}
37
+ {{- '' }}
38
+ {%- else %}
39
+ {{- raise_exception('Unexpected content type.') }}
40
+ {%- endif %}
41
+ {%- endmacro %}
42
+ {%- set ns_flags = namespace(enable_thinking=true) %}
43
+ {%- if enable_thinking is defined %}
44
+ {%- set ns_flags.enable_thinking = enable_thinking %}
45
+ {%- endif %}
46
+ {%- set preserve_thinking = preserve_thinking | default(true) %}
47
+ {%- if not messages %}
48
+ {{- raise_exception('No messages provided.') }}
49
+ {%- endif %}
50
+ {%- if add_generation_prompt is defined and add_generation_prompt and continue_final_message is defined and continue_final_message %}
51
+ {{- raise_exception('add_generation_prompt and continue_final_message cannot both be true.') }}
52
+ {%- endif %}
53
+ {%- if tools and tools is iterable and tools is not mapping %}
54
+ {{- '<|im_start|>system\n' }}
55
+ {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
56
+ {%- for tool in tools %}
57
+ {{- "\n" }}
58
+ {{- tool | tojson }}
59
+ {%- endfor %}
60
+ {{- "\n</tools>" }}
61
+ {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
62
+ {%- if messages[0].role == 'system' or messages[0].role == 'developer' %}
63
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
64
+ {%- if '<|think_off|>' in content %}
65
+ {%- set ns_flags.enable_thinking = false %}
66
+ {%- set content = content.replace('<|think_off|>', '') %}
67
+ {%- endif %}
68
+ {%- if '<|think_on|>' in content %}
69
+ {%- set ns_flags.enable_thinking = true %}
70
+ {%- set content = content.replace('<|think_on|>', '') %}
71
+ {%- endif %}
72
+ {%- set content = content.strip() %}
73
+ {%- if content %}
74
+ {{- '\n\n' + content }}
75
+ {%- endif %}
76
+ {%- endif %}
77
+ {{- '<|im_end|>\n' }}
78
+ {%- else %}
79
+ {%- if messages[0].role == 'system' or messages[0].role == 'developer' %}
80
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
81
+ {%- if '<|think_off|>' in content %}
82
+ {%- set ns_flags.enable_thinking = false %}
83
+ {%- set content = content.replace('<|think_off|>', '') %}
84
+ {%- endif %}
85
+ {%- if '<|think_on|>' in content %}
86
+ {%- set ns_flags.enable_thinking = true %}
87
+ {%- set content = content.replace('<|think_on|>', '') %}
88
+ {%- endif %}
89
+ {%- set content = content.strip() %}
90
+ {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
91
+ {%- endif %}
92
+ {%- endif %}
93
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
94
+ {%- for message in messages[::-1] %}
95
+ {%- set index = (messages|length - 1) - loop.index0 %}
96
+ {%- if ns.multi_step_tool and message.role == "user" %}
97
+ {%- set content = render_content(message.content, false)|trim %}
98
+ {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
99
+ {%- set ns.multi_step_tool = false %}
100
+ {%- set ns.last_query_index = index %}
101
+ {%- endif %}
102
+ {%- endif %}
103
+ {%- endfor %}
104
+ {%- if ns.multi_step_tool %}
105
+ {%- set ns.last_query_index = messages|length - 1 %}
106
+ {%- endif %}
107
+ {%- for message in messages %}
108
+ {%- set content = render_content(message.content, true)|trim %}
109
+ {%- set content = content.replace('<|think_off|>', '').replace('<|think_on|>', '') %}
110
+ {%- set content = content.strip() %}
111
+ {%- if message.role == "system" or message.role == "developer" %}
112
+ {%- if not loop.first %}
113
+ {%- set sys_content = render_content(message.content, false, true)|trim %}
114
+ {%- set sys_content = sys_content.replace('<|think_off|>', '').replace('<|think_on|>', '')|trim %}
115
+ {{- '<|im_start|>system\n' + sys_content + '<|im_end|>' + '\n' }}
116
+ {%- endif %}
117
+ {%- elif message.role == "user" %}
118
+ {{- '<|im_start|>' + message.role + '\n' + (content if content else ' ') + '<|im_end|>' + '\n' }}
119
+ {%- elif message.role == "assistant" %}
120
+ {%- set reasoning_content = '' %}
121
+ {%- if message.reasoning_content is string %}
122
+ {%- set reasoning_content = message.reasoning_content %}
123
+ {%- else %}
124
+ {%- set has_think_tag = false %}
125
+ {%- set think_start_token = '<think>' %}
126
+ {%- set think_end_token = '</think>' %}
127
+ {%- if '</think>' in content %}
128
+ {%- set has_think_tag = true %}
129
+ {%- elif '</thinking>' in content %}
130
+ {%- set has_think_tag = true %}
131
+ {%- set think_start_token = '<thinking>' %}
132
+ {%- set think_end_token = '</thinking>' %}
133
+ {%- elif '<think>' in content %}
134
+ {%- set reasoning_content = content.split('<think>')[-1].lstrip('\n') %}
135
+ {%- set content = '' %}
136
+ {%- elif '<thinking>' in content %}
137
+ {%- set reasoning_content = content.split('<thinking>')[-1].lstrip('\n') %}
138
+ {%- set content = '' %}
139
+ {%- endif %}
140
+ {%- if has_think_tag %}
141
+ {%- set reasoning_content = content.split(think_end_token)[0].rstrip('\n').split(think_start_token)[-1].lstrip('\n') %}
142
+ {%- set content = content.split(think_end_token)[-1].lstrip('\n') %}
143
+ {%- endif %}
144
+ {%- endif %}
145
+ {%- set reasoning_content = reasoning_content|trim %}
146
+ {%- set show_think = false %}
147
+ {%- if loop.index0 > ns.last_query_index and reasoning_content|length > 0 %}
148
+ {%- set show_think = true %}
149
+ {%- elif ns_flags.enable_thinking and preserve_thinking and reasoning_content|length > 0 %}
150
+ {%- set show_think = true %}
151
+ {%- endif %}
152
+ {%- if show_think %}
153
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
154
+ {%- else %}
155
+ {{- '<|im_start|>' + message.role + '\n' + content }}
156
+ {%- endif %}
157
+ {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
158
+ {%- for tool_call in message.tool_calls %}
159
+ {%- if tool_call.function is defined %}
160
+ {%- set tool_call = tool_call.function %}
161
+ {%- endif %}
162
+ {%- if loop.first %}
163
+ {%- if content|trim %}
164
+ {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
165
+ {%- else %}
166
+ {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
167
+ {%- endif %}
168
+ {%- else %}
169
+ {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
170
+ {%- endif %}
171
+ {%- if tool_call.arguments is defined and tool_call.arguments is mapping %}
172
+ {%- if tool_call.arguments|length > 0 %}
173
+ {%- for args_name in tool_call.arguments %}
174
+ {%- set args_value = tool_call.arguments[args_name] %}
175
+ {{- '<parameter=' + args_name + '>\n' }}
176
+ {%- set args_value = args_value | string if args_value is string else args_value | tojson %}
177
+ {{- args_value }}
178
+ {{- '\n</parameter>\n' }}
179
+ {%- endfor %}
180
+ {%- endif %}
181
+ {%- elif tool_call.arguments is defined and tool_call.arguments is string %}
182
+ {%- if tool_call.arguments|trim|length > 0 %}
183
+ {#- Note: raw JSON string arguments are emitted as-is and will not match
184
+ the XML parameter format in the tool instructions. Normalize arguments
185
+ to a dict in your serving layer before applying this template. -#}
186
+ {{- tool_call.arguments }}
187
+ {{- '\n' }}
188
+ {%- endif %}
189
+ {%- endif %}
190
+ {{- '</function>\n</tool_call>' }}
191
+ {%- endfor %}
192
+ {%- endif %}
193
+ {%- if not (loop.last and continue_final_message is defined and continue_final_message is true) %}
194
+ {{- '<|im_end|>\n' }}
195
+ {%- endif %}
196
+ {%- elif message.role == "tool" %}
197
+ {%- if not loop.previtem or (loop.previtem.role != "tool" and loop.previtem.role != "assistant") %}
198
+ {{- raise_exception('A tool message must follow an assistant or tool message.') }}
199
+ {%- endif %}
200
+ {%- if loop.previtem and loop.previtem.role != "tool" %}
201
+ {{- '<|im_start|>user' }}
202
+ {%- endif %}
203
+ {{- '\n<tool_response>\n' }}
204
+ {{- content }}
205
+ {{- '\n</tool_response>' }}
206
+ {%- if not loop.last and loop.nextitem.role != "tool" %}
207
+ {{- '<|im_end|>\n' }}
208
+ {%- elif loop.last %}
209
+ {{- '<|im_end|>\n' }}
210
+ {%- endif %}
211
+ {%- else %}
212
+ {{- raise_exception('Unexpected message role.') }}
213
+ {%- endif %}
214
+ {%- endfor %}
215
+ {%- if add_generation_prompt %}
216
+ {{- '<|im_start|>assistant\n' }}
217
+ {%- if ns_flags.enable_thinking is false %}
218
+ {{- '<think>\n\n</think>\n\n' }}
219
+ {%- else %}
220
+ {{- '<think>\n' }}
221
+ {%- endif %}
222
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3_5ForConditionalGeneration"
4
+ ],
5
+ "dtype": "bfloat16",
6
+ "image_token_id": 248056,
7
+ "language_model_only": false,
8
+ "model_type": "qwen3_5",
9
+ "text_config": {
10
+ "attention_bias": false,
11
+ "attention_dropout": 0.0,
12
+ "attn_output_gate": true,
13
+ "bos_token_id": 248044,
14
+ "dtype": "bfloat16",
15
+ "eos_token_id": 248044,
16
+ "full_attention_interval": 4,
17
+ "head_dim": 256,
18
+ "hidden_act": "silu",
19
+ "hidden_size": 5120,
20
+ "initializer_range": 0.02,
21
+ "intermediate_size": 17408,
22
+ "layer_types": [
23
+ "linear_attention",
24
+ "linear_attention",
25
+ "linear_attention",
26
+ "full_attention",
27
+ "linear_attention",
28
+ "linear_attention",
29
+ "linear_attention",
30
+ "full_attention",
31
+ "linear_attention",
32
+ "linear_attention",
33
+ "linear_attention",
34
+ "full_attention",
35
+ "linear_attention",
36
+ "linear_attention",
37
+ "linear_attention",
38
+ "full_attention",
39
+ "linear_attention",
40
+ "linear_attention",
41
+ "linear_attention",
42
+ "full_attention",
43
+ "linear_attention",
44
+ "linear_attention",
45
+ "linear_attention",
46
+ "full_attention",
47
+ "linear_attention",
48
+ "linear_attention",
49
+ "linear_attention",
50
+ "full_attention",
51
+ "linear_attention",
52
+ "linear_attention",
53
+ "linear_attention",
54
+ "full_attention",
55
+ "linear_attention",
56
+ "linear_attention",
57
+ "linear_attention",
58
+ "full_attention",
59
+ "linear_attention",
60
+ "linear_attention",
61
+ "linear_attention",
62
+ "full_attention",
63
+ "linear_attention",
64
+ "linear_attention",
65
+ "linear_attention",
66
+ "full_attention",
67
+ "linear_attention",
68
+ "linear_attention",
69
+ "linear_attention",
70
+ "full_attention",
71
+ "linear_attention",
72
+ "linear_attention",
73
+ "linear_attention",
74
+ "full_attention",
75
+ "linear_attention",
76
+ "linear_attention",
77
+ "linear_attention",
78
+ "full_attention",
79
+ "linear_attention",
80
+ "linear_attention",
81
+ "linear_attention",
82
+ "full_attention",
83
+ "linear_attention",
84
+ "linear_attention",
85
+ "linear_attention",
86
+ "full_attention"
87
+ ],
88
+ "linear_conv_kernel_dim": 4,
89
+ "linear_key_head_dim": 128,
90
+ "linear_num_key_heads": 16,
91
+ "linear_num_value_heads": 48,
92
+ "linear_value_head_dim": 128,
93
+ "mamba_ssm_dtype": "float32",
94
+ "max_position_embeddings": 262144,
95
+ "model_type": "qwen3_5_text",
96
+ "mtp_num_hidden_layers": 1,
97
+ "mtp_use_dedicated_embeddings": false,
98
+ "num_attention_heads": 24,
99
+ "num_hidden_layers": 64,
100
+ "num_key_value_heads": 4,
101
+ "output_gate_type": "swish",
102
+ "pad_token_id": null,
103
+ "partial_rotary_factor": 0.25,
104
+ "rms_norm_eps": 1e-06,
105
+ "rope_parameters": {
106
+ "mrope_interleaved": true,
107
+ "mrope_section": [
108
+ 11,
109
+ 11,
110
+ 10
111
+ ],
112
+ "partial_rotary_factor": 0.25,
113
+ "rope_theta": 10000000,
114
+ "rope_type": "default"
115
+ },
116
+ "tie_word_embeddings": false,
117
+ "use_cache": true,
118
+ "vocab_size": 248320
119
+ },
120
+ "tie_word_embeddings": false,
121
+ "transformers_version": "5.8.0",
122
+ "video_token_id": 248057,
123
+ "vision_config": {
124
+ "deepstack_visual_indexes": [],
125
+ "depth": 27,
126
+ "dtype": "bfloat16",
127
+ "hidden_act": "gelu_pytorch_tanh",
128
+ "hidden_size": 1152,
129
+ "in_channels": 3,
130
+ "initializer_range": 0.02,
131
+ "intermediate_size": 4304,
132
+ "model_type": "qwen3_5_vision",
133
+ "num_heads": 16,
134
+ "num_position_embeddings": 2304,
135
+ "out_hidden_size": 5120,
136
+ "patch_size": 16,
137
+ "spatial_merge_size": 2,
138
+ "temporal_patch_size": 2
139
+ },
140
+ "vision_end_token_id": 248054,
141
+ "vision_start_token_id": 248053
142
+ }
generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 248044,
3
+ "do_sample": true,
4
+ "eos_token_id": [
5
+ 248046,
6
+ 248044
7
+ ],
8
+ "pad_token_id": 248044,
9
+ "temperature": 1.0,
10
+ "top_k": 20,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.8.0"
13
+ }
model-00001-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:994d372da0812952c05a3fb1093db5b1245e314385290892f0dce3883f4c2643
3
+ size 2542796928
model-00002-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b6df0e83a1bbc861f46261a10ce7dc125db3e3371c161b96bc6cc8c2ce4735f2
3
+ size 4842451920
model-00003-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3aa1818f9e1a538044b18fb4a827bb03864fd051e6aebb99cb18b20d836e2f06
3
+ size 4965227944
model-00004-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:50d10d1df8ad52842b16dcb4c6e3a7639e7abadc90ebb60ba9ab5f7453ebc872
3
+ size 4912819264
model-00005-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb6d0c2d43dc0e20c4d3fd7d0d1be0c1ba9c9b6500f756df8b6eab265727bc19
3
+ size 4986198544
model-00006-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:78d60443c31e157a2015311dec3ee544e285cddf6163e41dcf9dbb88e4877fb1
3
+ size 4912819320
model-00007-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e968bfb9ece0b64c2825655a298d570fab00ae8dda9f691166ffb347c07ef89f
3
+ size 4932703272
model-00008-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:53ed65db4ece2d7b13a1b0c23944c3db5cbc8d60bb41d16beaaf15f9456df86b
3
+ size 4966314576
model-00009-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bda0ddb42d5cfd994e2c1b1a11608b75f53e88f8c66a0a23bf7e74af91c722cd
3
+ size 4964162248
model-00010-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa3296821e1c27efdb83550308b8eabb5fee6b2beccfd2b63ac89acff7434990
3
+ size 4933789824
model-00011-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e5f10b85b119c2ea613b0012268df35b45ab2ec40237eb0940c2595d5e69f94
3
+ size 4965228032
model-00012-of-00012.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0044ec0c32e4cc9efa02c83a486c37f27d6c337343f96293216930e96c848198
3
+ size 2789094896
model-auxiliary.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:713b0fafacc94c9e541925872de3bcc3507cf5af73abbce525298467b8b4b10f
3
+ size 849400392
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
preprocessor_config.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": {
3
+ "longest_edge": 16777216,
4
+ "shortest_edge": 65536
5
+ },
6
+ "patch_size": 16,
7
+ "temporal_patch_size": 2,
8
+ "merge_size": 2,
9
+ "image_mean": [
10
+ 0.5,
11
+ 0.5,
12
+ 0.5
13
+ ],
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "processor_class": "Qwen3VLProcessor",
20
+ "image_processor_type": "Qwen2VLImageProcessorFast"
21
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f32ce20dc35f57a7f9ad1eac03525bd7d30f9df8cea6507e958279cc3657706
3
+ size 19989492
tokenizer_config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "audio_bos_token": "<|audio_start|>",
4
+ "audio_eos_token": "<|audio_end|>",
5
+ "audio_token": "<|audio_pad|>",
6
+ "backend": "tokenizers",
7
+ "bos_token": null,
8
+ "clean_up_tokenization_spaces": false,
9
+ "eos_token": "<|im_end|>",
10
+ "errors": "replace",
11
+ "image_token": "<|image_pad|>",
12
+ "is_local": true,
13
+ "local_files_only": false,
14
+ "model_max_length": 262144,
15
+ "model_specific_special_tokens": {
16
+ "audio_bos_token": "<|audio_start|>",
17
+ "audio_eos_token": "<|audio_end|>",
18
+ "audio_token": "<|audio_pad|>",
19
+ "image_token": "<|image_pad|>",
20
+ "video_token": "<|video_pad|>",
21
+ "vision_bos_token": "<|vision_start|>",
22
+ "vision_eos_token": "<|vision_end|>"
23
+ },
24
+ "pad_token": "<|endoftext|>",
25
+ "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
26
+ "split_special_tokens": false,
27
+ "tokenizer_class": "Qwen2Tokenizer",
28
+ "unk_token": null,
29
+ "video_token": "<|video_pad|>",
30
+ "vision_bos_token": "<|vision_start|>",
31
+ "vision_eos_token": "<|vision_end|>"
32
+ }
video_preprocessor_config.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": {
3
+ "longest_edge": 25165824,
4
+ "shortest_edge": 4096
5
+ },
6
+ "patch_size": 16,
7
+ "temporal_patch_size": 2,
8
+ "merge_size": 2,
9
+ "image_mean": [
10
+ 0.5,
11
+ 0.5,
12
+ 0.5
13
+ ],
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "processor_class": "Qwen3VLProcessor",
20
+ "video_processor_type": "Qwen3VLVideoProcessor"
21
+ }