--- library_name: mlx tags: - mlx - oq - quantized --- # Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp This model was quantized using [oQ](https://github.com/jundot/omlx) (oMLX v0.3.12) mixed-precision quantization. ## Quantization details - **Model type**: qwen3_5_moe - **Bits**: 5 - **Group size**: 64 - **Format**: MLX safetensors - ## Community Benchmark Results > Benchmarks run independently on local hardware using the [oMLX](https://github.com/jundot/omlx) inference server (Apple Silicon, MLX backend), all models with `think = yes`. > Timing data is omitted — inference was subject to swap memory pressure during these runs and is not a reliable proxy for reasoning depth or token budget. Accuracy is the signal. > Note: official accuracy benchmarks for the Claude-distilled base model were still marked as pending on the model card at time of writing. These runs represent some of the first independently verified results. ### Models | Tag | Model | Source | |-----|-------|--------| | **M1** | `Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ6e-fp16` | [splats/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ6e](https://huggingface.co/splats/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ6e) — oQ6e enhanced quantization of [lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled) | | **M2** | `Qwen3.6-35B-A3B-oQ6-fp16-mtp` | Independently quantized with [oMLX oQ](https://github.com/jundot/omlx) from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) | | **M3** | `Qwen3.6-35B-A3B-oQ5-fp16-mtp` | Independently quantized with [oMLX oQ](https://github.com/jundot/omlx) from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) | | **M4** | `Qwen3.6-35B-A3B-oQ4-fp16-mtp` | Independently quantized with [oMLX oQ](https://github.com/jundot/omlx) from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) | | **M5** | `Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4-fp16-mtp` | Independently quantized with [oMLX oQ](https://github.com/jundot/omlx) from [lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled) | ### Accuracy Bold = highest score for that benchmark across all models. Format: `accuracy% (correct)`. | Benchmark | Total | M1 Claude oQ6e | M2 base oQ6 | M3 base oQ5 | M4 base oQ4 | M5 Claude oQ4 | |-----------|------:|---------------:|------------:|------------:|------------:|--------------:| | MMLU | 1000 | 86.9% (869) | **90.0% (900)** | 82.4% (824) | 79.9% (799) | 66.3% (663) | | MMLU PRO | 300 | **83.3% (250)** | 83.0% (249) | 69.7% (209) | 68.7% (206) | 59.0% (177) | | HellaSwag | 200 | 92.5% (185) | **94.0% (188)** | 87.5% (175) | 87.0% (174) | 81.0% (162) | | TruthfulQA | 817 | 85.2% (696) | **90.1% (736)** | 82.1% (671) | 81.6% (667) | 66.7% (545) | | ARC Challenge | 300 | **97.0% (291)** | **97.0% (291)** | 86.0% (258) | 85.7% (257) | 73.3% (220) | | WinoGrande | 300 | 84.3% (253) | 78.3% (235) | **86.0% (258)** | 80.7% (242) | 60.0% (180) | | GSM8K | 100 | **95.0% (95)** | 85.0% (85) | 86.0% (86) | 85.0% (85) | 73.0% (73) | | MathQA | 300 | **91.3% (274)** | 77.3% (232) | 82.3% (247) | 76.0% (228) | 67.7% (203) | | HumanEval | 164 | **84.8% (139)** | 82.9% (136) | 82.9% (136) | 80.5% (132) | 69.5% (114) | | MBPP | 200 | 78.0% (156) | 81.5% (163) | **84.0% (168)** | 82.0% (164) | 68.0% (136) | | LiveCodeBench | 100 | 45.0% (45) | 42.0% (42) | **48.0% (48)** | 46.0% (46) | 34.0% (34) | | BBQ | 300 | **96.3% (289)** | 87.0% (261) | 88.3% (265) | 87.0% (261) | 85.7% (257) | | SafetyBench | 300 | **85.0% (255)** | 80.7% (242) | 81.3% (244) | 80.7% (242) | 75.0% (225) | ### Notes - **oQ6e distillation is worth it for math and reasoning.** M1 leads on 7 of 13 benchmarks, with the clearest gains on GSM8K (+10pp), MathQA (+9pp), and BBQ (+8pp). MMLU and TruthfulQA are the exceptions, where the base oQ6 edges ahead. - **oQ4 distillation is not recommended.** M5 scores lower than the base oQ4 on every benchmark without exception — an average drop of −10.8pp across all 13 tasks. The regression is not limited to coding; WinoGrande (commonsense reasoning) drops the most at −20.7pp. The distilled weights do not appear to survive 4-bit compression intact. The only near-parity is BBQ (−1.3pp), suggesting safety-adjacent behaviour is the most compression-resilient aspect of the distillation. - **Coding is the weakest area for the distilled model at any bit level.** Both M1 and M5 underperform on MBPP and LiveCodeBench relative to base models at the same or lower bit level. This is consistent with a [known limitation noted by the base model author](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled): the distillation used attention-only LoRA (0.01% of parameters), leaving the MoE expert FFNs unchanged — *"distillation transfers how to reason, not new facts."* --- *Benchmarks run independently; not affiliated with any model author. Results reflect a single hardware configuration and may differ on other setups.* ## Citation If you use these quantizations, please cite the base model and the distillation: ```bibtex @misc{qwen36_a3b_2026, title = {Qwen3.6-35B-A3B}, author = {Qwen Team}, year = {2026}, howpublished = {\url{https://huggingface.co/Qwen/Qwen3.6-35B-A3B}}, } @misc{lordx64_qwen36_distill_2026, title = {Qwen3.6-35B-A3B distilled from Claude Opus 4.7 reasoning}, author = {lordx64}, year = {2026}, howpublished = {\url{https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled}}, } ```