--- license: other language: - en - zh tags: - qwen3.6 - dsv4pro - glm - sft - rl - coding - gguf - fp8 - mtp base_model: Merkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill --- # Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding ![Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding](assets/model_card_hero_hf_lcb_20260714.png) > **Loadable GGUF files (2026-07-12)** > - Q3, Q4, and Q5 now provide directly loadable single-file main GGUFs with `SHA256SUMS` and `manifest.json`. > - Q8 remains a valid llama.cpp standard split; select `00001-of-00005.gguf` and the remaining shards load automatically. > - `*-draft*.gguf` files are optional MTP sidecars, not main models. LM Studio and general GUI launchers should select only `GGUF//Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding--LynnStyle.gguf`. Main Hugging Face repository for `Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding`. BF16 weights live at the repository root. GGUF artifacts are mirrored under `GGUF/`, and FP8 artifacts are mirrored under `FP8/`. Standalone GGUF repository: [nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF](https://huggingface.co/nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF). This model is the coding-focused post-training branch of `Qwen3.6-27B-DSV4Pro-Thinking-Distill`. The goal is not to make the model think longer. The goal is to make it deliver more reliably: write executable code, repair from real observations, verify before closing, and keep MMLU / GPQA / LCB / Coding100 within healthy gates. The standalone GGUF repository carries the same GGUF descriptions: [nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF](https://huggingface.co/nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF). This main repository carries BF16, GGUF, and FP8 artifacts. ## Which Directory Should I Download? - `GGUF/Q3_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q3-LynnStyle.gguf`: Q3 low-memory tier, complete single file; GGUF parsing plus Spark bare/MTP load and generation smoke passed. - `GGUF/Q3_LynnStyle/Q3-imatrix-MTP-draft.gguf`: optional Q3 MTP sidecar for llama.cpp `--model-draft` only. - `GGUF/Q4_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q4-LynnStyle.gguf`: Q4 24GB mainstream tier, complete single file. - `GGUF/Q4_LynnStyle/Q4-imatrix-MTP-draft.gguf`: Q6_K high-acceptance MTP sidecar; `Q4-imatrix-MTP-draft-q4.gguf` is the lower-memory alternative. - `GGUF/Q5_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q5-LynnStyle.gguf`: Q5 32GB high-quality tier, complete single file. - `GGUF/Q5_LynnStyle/Q5-imatrix-MTP-draft.gguf`: optional Q5 MTP sidecar. - `GGUF/Q8_0/`: valid standard-split Q8; load `00001-of-00005.gguf`. LM Studio and general GUI launchers should select only a complete main file, never an MTP sidecar as the main model. ## llama.cpp Quick Start General GUI clients should load only the main model. Q3 bare example: ```bash llama-server \ -m GGUF/Q3_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q3-LynnStyle.gguf \ -ngl 999 -c 8192 -b 1024 -ub 256 -np 1 \ --host 0.0.0.0 --port 8080 \ --jinja --reasoning on --reasoning-format deepseek ``` Q3 single-stream MTP example: ```bash llama-server \ -m GGUF/Q3_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q3-LynnStyle.gguf \ --model-draft GGUF/Q3_LynnStyle/Q3-imatrix-MTP-draft.gguf \ -ngl 999 --split-mode layer \ -c 8192 -b 1024 -ub 256 -t 16 -np 4 -cb \ --spec-type draft-mtp \ --spec-draft-n-max 3 \ --spec-draft-p-min 0.6 \ --host 0.0.0.0 --port 8080 \ --jinja --reasoning on --reasoning-format deepseek ``` Recommended: Q3 C1 uses MTP `n=3`; Q4 C1/C2 use `n=2`; Q5 C1 uses `n=3` and C2 uses `n=4`. Prefer bare for C4 throughput, long context, or multi-concurrency. ## Current Tiers | Memory | Tier | Status | Notes | |---:|---|---|---| | 16GB | Q3 LynnStyle Dense | **Verified single file** | On a strict 16GB setup, start with the main model only and tune context to available memory; MTP needs extra headroom. | | 24GB | Q4 LynnStyle Dense | **Verified single file** | Load the complete main GGUF directly in LM Studio or general launchers. | | 32GB | Q5 LynnStyle Dense | **Verified single file** | Complete main GGUF and optional MTP are published. | | 48GB+ | Q8_0 | Published | Valid llama.cpp standard split remains available. | ## Why LynnStyle Dense? MoE models can be inspected through expert heatmaps: which experts are active for MMLU / GPQA, LCB / Coding, and ReAct-style tasks. This 27B model is Dense, so there are no MoE experts to prune or protect, but the model is still not homogeneous. The protection unit changes from MoE's `layer + expert + tensor family` to Dense's: - `layer position` - `attention pattern` - `tensor family` Qwen3.6 27B uses mixed attention patterns. Full-attention layers appear roughly every four layers and are important for long-range integration and final-answer closure. Other layers rely more on linear-attention / SSM paths. The middle of the network is not an unprotected blank zone: full-attention projections, attention gates, linear-attention qkv, and SSM tensor families are all protected work areas. The GGUF probe confirms that the protection rules hit real tensors: | Protected group | Matched tensors | |---|---:| | early layers | 47 | | last four layers | 47 | | L56-L59 shoulder | 47 | | middle full-attention projections | 52 | | middle linear-attention gates | 39 | | middle linear-attention qkv | 39 | | middle SSM tensor family | 195 | For Q5, LynnStyle means explicit protected tensors plus imatrix calibration: | Region | Explicit precision | Matched tensors | |---|---|---:| | L0-L3 early layers | `q8_0` hard protection | 47 | | L60-L63 last layers | `q8_0` hard protection | 47 | | L56-L59 shoulder | `q6_k` protection | 47 | | middle full-attention q/k/v/o | `q5_k` protection | 52 | | middle linear-attention gate | `q5_k` protection | 39 | | middle linear-attention qkv | `q5_k` protection | 39 | | middle SSM tensor family | `q5_k` protection | 195 | Direct quantization alone is not a release tier. A LynnStyle Dense release tier must combine explicit core-layer / tensor-family protection with imatrix or equivalent importance calibration, and then pass public gates. ## Current Q8 Gate > Abbreviation: `LCB` stands for `LiveCodeBench`. Evaluation uses Q8_0 GGUF with the official MTP head, llama.cpp serving, temperature `0.6`, top_p `0.95`. | Dimension | Metric | Score | Meaning | |---|---|---:|---| | General knowledge | MMLU500 | 462/500 = 92.4% | Multi-domain knowledge and understanding. | | Hard reasoning | GPQA198 | 162/198 = 81.82% | Scientific reasoning and final-answer closure. | | Algorithmic code | LCB100 | 78/100 | Hidden-test coding and algorithmic problem solving. | | Multilingual code | Coding100 clean | 88/100 | Multi-language engineering coverage. | Coding100 language split: Python 9/10, JavaScript 6/10, TypeScript 7/10, Rust 9/10, Go 8/10, C++ 10/10, SQL 10/10, Bash 9/10, CSS 10/10, HTML 10/10. LCB100 difficulty split: easy 23/23, medium 30/31, hard 25/46. ## Cross-Version Clean Scores Only complete clean results are listed here. Harness-polluted rows are retested by id and merged before a score is published. | Version | MMLU500 | GPQA198 | LCB100 | Coding100 | |---|---:|---:|---:|---:| | Qwen3.6-27B original Q8 | 91.6% | 73.7% | 68/100 | 83/100 | | DSV4Pro distill Q8 | 91.8% | 80.81% | 74/100 | 86/100 | | This release Q8 | 92.4% | 81.82% | 78/100 | 88/100 | | Q5 LynnStyle | 92.2% | 78.79% | 75/100 | 81/100 | | Q4 LynnStyle | 92.0% | 74.75% | 71/100 | 82/100 | | Q3 LynnStyle | 91.4% | 77.27% | 69/100 | 84/100 | Original Q8 LCB100 clean result: `68/100`; difficulty split: easy `23/23`, medium `27/31`, hard `18/46`. Three JSONDecodeError-polluted rows were retested non-streaming by exact id and all three passed, leaving `errors=0`. DSV4Pro distill Q8 LCB100 clean result: `74/100`; difficulty split: easy `23/23`, medium `25/31`, hard `26/46`; `errors=0`. ## Q5 LynnStyle Gate Q5 is the 32GB recommended tier. It is designed to stay close to Q8 quality while reducing file size enough for local deployment. | Metric | Q5 LynnStyle | Status | |---|---:|---| | MMLU500 | 461/500 = 92.2% | clean | | GPQA198 | 156/198 = 78.79% | clean | | LCB100 | 75/100 | clean | | Coding100 current clean | 81/100 | clean | Q5 Coding100 current clean result: 81/100; language split: Python 9/10, JavaScript 5/10, TypeScript 8/10, Rust 8/10, Go 5/10, C++ 8/10, SQL 9/10, Bash 9/10, CSS 10/10, HTML 10/10. Three timeout ids were retested by exact id and merged; API errors after retest: 0. ## Q4 LynnStyle Gate Q4 is the 24GB mainstream tier. For 16K/24K long-context use on 24GB GPUs, load the Q4 main model only and do not load the MTP sidecar. The Q4 MTP sidecar is optional for short-context or lower-concurrency speculative decoding when VRAM has headroom. | Metric | Q4 LynnStyle | Status | |---|---:|---| | MMLU500 | 460/500 = 92.0% | clean | | GPQA198 | 148/198 = 74.75% | clean | | LCB100 | 71/100 | clean | | Coding100 current clean | 82/100 | clean | The GPQA198 result was merged from the interrupted first pass and a missing-id resume pass while preserving the original GPQA ids and choice mappings. Health counters: `dirty=0`, `parse_fail=0`, `empty_prediction=0`, `duplicate_ids=0`; finish reasons: `stop=197`, `length=1`. Q4 LCB100 full clean result: `71/100`; difficulty split: easy `23/23`, medium `27/31`, hard `21/46`. Failure notes: wrong answer `20`, runtime error `4`, timeout `5`; no harness error/jsondecode/parse pollution. Q4 Coding100 current clean result: `82/100`; language split: Python `9/10`, JavaScript `6/10`, TypeScript `9/10`, Rust `7/10`, Go `6/10`, C++ `7/10`, SQL `10/10`, Bash `8/10`, CSS `10/10`, HTML `10/10`. API errors: `0`. ## Q4 / Q3 llama.cpp Runtime Matrix TPS is rounded to whole numbers and acceptance is shown as a percentage. C1/C2/C4 denote concurrency; all rows use the same llama.cpp harness. **Q4 LynnStyle** | Mode | C1 | C2 | C4 | |---|---:|---:|---:| | bare | 59 TPS | 76 TPS | **116 TPS** | | best MTP | n=2: 75 TPS (91%) | n=2: 82 TPS (93%) | n=3: 83 TPS (89%) | Best tier setting: **C4 bare, 116 TPS**. For lower-concurrency latency, use MTP `n=2` at C1/C2. **Q3 LynnStyle** | Concurrency | bare | MTP n=2 | MTP n=3 | MTP n=4 | Recommendation | |---|---:|---:|---:|---:|---| | C1 | 15 TPS | 22 TPS (86%) | 23 TPS (84%) | 23 TPS (81%) | **MTP n=3**, about +53% vs bare | | C2 | 28 TPS | 28 TPS (86%) | 28 TPS (84%) | 27 TPS (81%) | **MTP n=2**, tied with bare after rounding | | C4 | 49 TPS | 29 TPS (86%) | 30 TPS (84%) | 30 TPS (81%) | **bare** | For Q3 single-request latency, use MTP `n=3`. At C2, the best MTP choice is `n=2`, but rounded throughput ties bare. For C4 throughput, use bare. On strict 16GB or long-context setups, load only the Q3 main model first. ## Training and Data Hygiene This release does not train on benchmark answer keys. SFT teaches trajectories; RL teaches behavior preferences. Benchmarks remain held out to check whether the model learned to think, solve, act, verify, and close. High-level training route: | Stage | Purpose | Data shape | |---|---|---| | SFT-1A direct-code | Improve executable code and weak programming languages. | Direct code / single-code-block tasks. | | SFT-1B real-observation ReAct | Teach reason -> act -> observe -> repair -> verify -> final. | Real observation trajectories. | | RL-1 clean | Reduce empty answers, fake verification, no-final, and invalid loops. | Preference pairs. | | RL-2 GPT55 guarded | Final lightweight behavior calibration. | Focused preference pairs for final-answer closure, anti-insufficient behavior, execution repair, and reasoning-budget control. | Data policy: - GLM-5.2 is used for candidate generation and coverage. - GPT55 / Codex-style review is used for judging, correction, and high-value preference pairs. - Real execution observations have priority as evidence. - Tool observations are used as context, not as assistant loss targets. - Assistant-only loss trains assistant thought / action / repair / final segments. ## 中文说明 > **GGUF 可加载文件（2026-07-12）** > - Q3、Q4、Q5 均已发布可直接加载的单文件主 GGUF，并附 `SHA256SUMS` 与 `manifest.json`。 > - Q8 保持 llama.cpp 标准多分片；选择 `00001-of-00005.gguf` 后会自动加载其余分片。 > - `*-draft*.gguf` 是可选 MTP sidecar，不是主模型；LM Studio 与通用图形启动器只选择 `GGUF/<档位目录>/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-<档位>-LynnStyle.gguf`。 这是 `Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding` 的 Hugging Face 主仓。根目录放 BF16 权重；`GGUF/` 放从独立 GGUF 仓同步的量化文件、MTP sidecar、runtime smoke、MTP/TPS 矩阵和量化门禁结果；`FP8/` 放 FP8 权重、SGLang 启动脚本和 MTP/NEXTN TPS 数据。 这不是“让模型更啰嗦思考”的版本，而是把 DSV4Pro 蒸馏版继续往 coding / execution / 可交付方向推：能写可执行代码，能基于真实 observation 修复，完成前会验证，收口时给出明确答案，同时尽量守住 MMLU / GPQA / LCB100 / Coding100。 ### 下载目录 - `GGUF/Q3_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q3-LynnStyle.gguf`：Q3 低显存档，完整单文件；GGUF 解析与 Spark bare/MTP 加载生成 smoke 均通过。 - `GGUF/Q3_LynnStyle/Q3-imatrix-MTP-draft.gguf`：Q3 可选 MTP sidecar，仅供 llama.cpp `--model-draft`。 - `GGUF/Q4_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q4-LynnStyle.gguf`：Q4 24GB 主流档，完整单文件。 - `GGUF/Q4_LynnStyle/Q4-imatrix-MTP-draft.gguf`：Q6_K 高接受率 MTP sidecar；`Q4-imatrix-MTP-draft-q4.gguf` 是低显存备选。 - `GGUF/Q5_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q5-LynnStyle.gguf`：Q5 32GB 高质量档，完整单文件。 - `GGUF/Q5_LynnStyle/Q5-imatrix-MTP-draft.gguf`：Q5 可选 MTP sidecar。 - `GGUF/Q8_0/`：标准 GGUF 多分片 Q8；加载 `00001-of-00005.gguf`。 LM Studio、llama.cpp Windows Manager 等图形启动器只选择完整主文件，不要把 MTP sidecar 当主模型。 ### llama.cpp 启动 通用图形客户端只加载主模型。Q3 bare 示例： ```bash llama-server \ -m GGUF/Q3_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q3-LynnStyle.gguf \ -ngl 999 -c 8192 -b 1024 -ub 256 -np 1 \ --host 0.0.0.0 --port 8080 \ --jinja --reasoning on --reasoning-format deepseek ``` Q3 单流 MTP 示例： ```bash llama-server \ -m GGUF/Q3_LynnStyle/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-Q3-LynnStyle.gguf \ --model-draft GGUF/Q3_LynnStyle/Q3-imatrix-MTP-draft.gguf \ -ngl 999 --split-mode layer \ -c 8192 -b 1024 -ub 256 -t 16 -np 4 -cb \ --spec-type draft-mtp \ --spec-draft-n-max 3 \ --spec-draft-p-min 0.6 \ --host 0.0.0.0 --port 8080 \ --jinja --reasoning on --reasoning-format deepseek ``` 推荐：Q3 C1 用 MTP `n=3`；Q4 C1/C2 用 `n=2`；Q5 C1 用 `n=3`、C2 用 `n=4`。C4 吞吐优先使用 bare。长上下文或多并发也优先 bare。 ### LynnStyle Dense 27B 是 Dense，没有 MoE expert，但不代表全层同质。Qwen3.6 27B 有混合注意力结构：full-attention 层约每 4 层出现一次，其他层更多是 linear attention / SSM。LynnStyle Dense 的保护单位是层位、attention pattern、tensor family。 Q5/Q4/Q3/Q2 不是一刀切低比特量化，而是显式保护关键层、full-attention、attention gate、SSM 和重要 tensor family，再对普通层按档位压缩，并叠加 imatrix 或等价重要度校准。只做 direct quant 不具备发布资格。 ### 当前门禁 - Q8：MMLU500 92.4%，GPQA198 81.82%，LCB100 78/100，Coding100 clean 88/100。 - Q5：MMLU500 92.2%，GPQA198 78.79%，LCB100 75/100，Coding100 clean 81/100。 - Q4：MMLU500 92.0%，GPQA198 74.75%，LCB100 71/100，Coding100 clean 82/100。 - Q3：MMLU500 91.4%，GPQA198 77.27%，LCB100 69/100，Coding100 clean 84/100。 训练不灌 benchmark 答案。SFT 学轨迹，RL 学行为偏好，MMLU / GPQA / LCB / Coding100 作为 held-out 门禁。 ## Q4 / Q3 llama.cpp 运行矩阵 TPS 取整，接受率以百分比显示。C1/C2/C4 表示并发数；结果来自同一套 llama.cpp harness。 **Q4 LynnStyle** | 模式 | C1 | C2 | C4 | |---|---:|---:|---:| | bare | 59 TPS | 76 TPS | **116 TPS** | | 最佳 MTP | n=2：75 TPS（91%） | n=2：82 TPS（93%） | n=3：83 TPS（89%） | 最佳档位：**C4 bare，116 TPS**。低并发延迟优先时，C1/C2 使用 MTP `n=2`。 **Q3 LynnStyle** | 并发 | bare | MTP n=2 | MTP n=3 | MTP n=4 | 推荐 | |---|---:|---:|---:|---:|---| | C1 | 15 TPS | 22 TPS（86%） | 23 TPS（84%） | 23 TPS（81%） | **MTP n=3**，较 bare 约 +53% | | C2 | 28 TPS | 28 TPS（86%） | 28 TPS（84%） | 27 TPS（81%） | **MTP n=2**，取整后与 bare 同速 | | C4 | 49 TPS | 29 TPS（86%） | 30 TPS（84%） | 30 TPS（81%） | **bare** | Q3 单请求延迟优先推荐 MTP `n=3`；C2 的 MTP 最优选择为 `n=2`，但取整吞吐与 bare 持平；C4 吞吐优先使用 bare。严格 16GB 或长上下文也优先只加载 Q3 主模型。