Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding

Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding

<!-- Q4-SINGLE-FILE-EN-START -->

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/<tier directory>/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-<tier>-LynnStyle.gguf.

<!-- Q4-SINGLE-FILE-EN-END -->

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.

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. 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:

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:

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

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:

For Q5, LynnStyle means explicit protected tensors plus imatrix calibration:

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

<!-- LCB-ABBREVIATION-EN-START -->

Abbreviation: LCB stands for LiveCodeBench.

<!-- LCB-ABBREVIATION-EN-END -->

Evaluation uses Q8_0 GGUF with the official MTP head, llama.cpp serving, temperature 0.6, top_p 0.95.

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.

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.

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.

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.

<!-- GGUF-RUNTIME-MATRIX-EN-START -->

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

Best tier setting: C4 bare, 116 TPS. For lower-concurrency latency, use MTP n=2 at C1/C2.

Q3 LynnStyle

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.

<!-- GGUF-RUNTIME-MATRIX-EN-END -->

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:

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

<!-- Q4-SINGLE-FILE-ZH-START -->

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。

<!-- Q4-SINGLE-FILE-ZH-END -->

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：

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：

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。

<!-- GGUF-RUNTIME-MATRIX-ZH-START -->

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

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

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。

<!-- GGUF-RUNTIME-MATRIX-ZH-END -->