license: apache-2.0
language:

• en
• zh
  tags:
• qwen3.8
• gguf
• efficient-thinking
• dflash2
• speculative-decoding
• mtp
• llama.cpp
• q2-lynnstyle
• mixed-precision
• gsq-rco
• gsq
• iq
• rco

Qwen3.8-27B EfficientThink GGUF

No strict loops were observed in the reviewed Q2–Q8 evaluations.

BF16 / FP8 main repository: nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2

Measured quantization-tier overview

Capability scores are shown only after a formal full-suite freeze, with every non-passing sample retained in the denominator. Q8_0, Q6_K, Q5-LynnStyle, Q4-LynnStyle, Q3-LynnStyle, and Q2-LynnStyle are published and have passed four-repository remote path, byte-size, and LFS-SHA256 readback.

Lynn Agent v0.87.0

Lynn Agent v0.87.0 uses this release's Q2-LynnStyle / Q3-LynnStyle + DFlash2 packages. The pairing passed runtime validation on DGX Spark; notarized Mac Apple Silicon and Intel builds, the Windows installer runtime check, CI in both repositories, matching main heads and tags across all three release repositories, complete SHA256 verification for 23 public files, and a remote CLI installation all passed.

Natural moving tree shadows and soft window light, enabled by default. Hover over Shadows for the off switch location, or click to open Settings. Playback pauses in the background and stays still with reduced motion. Images now participate in file filtering; slash templates replace persistent task mode; translation moved into the message menu; Expert Roundtable is now an optional plugin; and session edit targeting and stop preprocessing were fixed. Kimi Datasource remains available under MCP and requires users to scan the QR code and sign in with their own account.

This client update did not change any model weight, quantized artifact, benchmark score, or performance metric in this repository.

Release records: primary GitHub repository · legacy GitHub repository · Gitee · CLI package

llama.cpp prerequisites and launch settings

Use an official llama.cpp build containing DFlash2 support, merged 2026-08-27, commit b10f9ca58c89 or a compatible newer build. A reproducible minimum build from the official repository is:

git clone https://github.com/ggml-org/llama.cpp.git llama.cpp-efficientthink
git -C llama.cpp-efficientthink checkout b10f9ca58c89ccfc3653ac01e979dd085d582b76
cmake -S llama.cpp-efficientthink -B llama.cpp-efficientthink/build -DGGML_CUDA=ON
cmake --build llama.cpp-efficientthink/build --config Release -j 4 --target llama-server
export PATH="$PWD/llama.cpp-efficientthink/build/bin:$PATH"
llama-server --version
llama-server --help | grep -- draft-dflash

Run the tier commands below from this repository's download root. If using the main repository's mirror, first cd GGUF. Choose one tier/server at a time. Each command explicitly selects GPU 0; remove or adjust CUDA_VISIBLE_DEVICES=0 only for your own multi-GPU configuration. CUDA Toolkit and a compiler supported by the chosen GPU are prerequisites; the Spark FP8 Docker image is not required for GGUF.

The measured layout is 8 slots × 32,768 tokens = --ctx-size 262144 total, not 32K total shared by eight slots. C4/C8 means four/eight active client requests against the eight-slot server. On smaller-memory hardware reduce both slots and total context (for example one slot with --parallel 1 --ctx-size 32768); those settings do not inherit the published C4/C8 throughput. Weight size alone is not the total runtime memory requirement.

DFlash2 can use either the separate Q8_0 or Q4_K_M draft, --spec-type draft-dflash, --spec-draft-n-max 7, and --spec-draft-n-min 0. It is not the in-model MTP path. --jinja and explicit XH template arguments retain the model's thinking mode. The examples bind to 127.0.0.1:8080; expose a different interface only with appropriate access controls. The --mmproj sidecar is for images; omit that line for text-only serving.

Official references: server options, speculative decoding, CUDA build. These corrected commands match the documented flags and measured slot layout; the build instructions are not a new six-tier benchmark run.

Why DFlash2, and bundled MTP alternatives

DFlash2 remains our preferred speculative decoder for its measured draft acceptance and output TPS. Parallel block drafting and acceptance of multiple tokens per verification help reduce sequential decoding overhead. Method reference: DFlash authors.

For example, the measured Q8_0 C4 configuration achieved 60.98% acceptance and 246 aggregate tok/s; Q3-LynnStyle C4 achieved 58.00% and 269 aggregate tok/s. The tables below report each tier independently: C4 is the practical balance and C8 the maximum measured throughput. Acceptance and TPS depend on the model, hardware, workload, and concurrency; these results are not a matched MTP comparison.

Q4_0 and Q8_0 MTP GGUF files are bundled

Every Q2-LynnStyle through Q8_0 directory now contains the same two independently verified MTP sidecars:

Both are derived from ggml-org/Qwen3.8-27B-GGUF@0669b98607d47046c7c2b3f801011d54a08cfccf and parse as Qwen MTP GGUFs with 65 blocks and 18 tensors. A real DGX Spark smoke paired each file with the published Q3-LynnStyle main model: both loaded, returned HTTP 200, generated non-empty final content, and stopped normally. This is a compatibility smoke, not a frozen MTP TPS or capability benchmark.

Choose one speculative path per launch. Do not combine DFlash2 and MTP. Example with the smaller MTP draft:

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q3-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q3-LynnStyle.gguf \
  --model-draft Q3-LynnStyle/mtp-Qwen3.8-27B-Q4_0.gguf \
  --spec-type draft-mtp \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 1 --ctx-size 32768 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Substitute mtp-Qwen3.8-27B-Q8_0.gguf for the Q8_0 MTP option. Add the tier's --mmproj only when serving images.

Per-tier DFlash2 concurrency performance

Common measured protocol: one GPU for the main model, DFlash2, --parallel 8 --ctx-size 262144 (8 slots × 32K), C1/C2/C4/C8, fixed 256-token pressure outputs, and a 60-second request ceiling. DFlash2 used n_max=7 and n_min=0. Throughput is aggregate completion tok/s.

Q8_0

• Concurrency guidance: C4 practical balance; C8 maximum throughput.
• 15 fixed-length requests per tier; zero timeout, request error, empty output, or punctuation collapse.

Q6_K

• Concurrency guidance: C4 lower-latency balance; C8 maximum throughput.
• Zero timeout, request error, empty output, or punctuation collapse across the sweep.

Q5-LynnStyle

• Concurrency guidance: C4 practical balance; C8 maximum throughput.
• Zero timeout, request error, empty output, or punctuation collapse across the sweep.

Q4-LynnStyle

• Concurrency guidance: C4 practical balance; C8 maximum throughput.
• Zero timeout, request error, empty output, or punctuation collapse across the sweep.

Q3-LynnStyle

• Concurrency guidance: C4 is the practical balance; C8 is the maximum measured throughput.
• The full sweep had zero timeout, request error, empty output, or punctuation collapse.

Training method

Qwen/Qwen3.8-27B → capability-preserving SFT → terminal-behavior SimPO → per-tensor FP32 delta merge → BF16 → GGUF quantization.

• SFT: 1,905 samples, 1 epoch, 239 optimizer steps, effective batch 8; LoRA r=16, alpha=32, dropout=0.05; LR 5e-6, 12 warmup steps, seed 20260901.
• SimPO: 110 preference pairs over 73 unique prompts, 5 optimizer steps; beta=1.0, gamma=0.2, peak LR 5e-7; LoRA r=16, alpha=32, dropout=0, seed 20260903, world size 2, FSDP full sharding.
• Training hardware: 2x NVIDIA RTX PRO 6000 Blackwell Server Edition.

This repository publishes text-main GGUF files with independently verified vision and DFlash2 sidecars. Q8_0 and Q6_K use the Q8 vision projector; Q5-LynnStyle, Q4-LynnStyle, and Q3-LynnStyle use the Q6 vision projector. All six tiers provide Q8_0 and smaller Q4_K_M DFlash2 drafts plus Q4_0 and Q8_0 MTP sidecars. Choose exactly one speculative path per launch; none of these sidecars is a main model.

Published GGUF files

Q8_0

Formal full-suite evaluation: GPQA 164/198 (82.83%) · MMLU 447/500 (89.40%) · LCB 74/100 (74.00%).

Concurrency guidance: C4 is the practical balance (246 tok/s, 60.98% acceptance); C8 is the maximum measured throughput (256 tok/s, 52.01%).

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q8_0/Qwen3.8-27B-EfficientThink-SimPO-Q8_0.gguf \
  --mmproj Q8_0/mmproj-Qwen3.8-27B-Q8_0.gguf \
  --model-draft Q8_0/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q5-LynnStyle

Formal full-suite evaluation: GPQA 164/198 (82.83%) · MMLU 438/500 (87.60%) · LCB 75/100 (75.00%). LCB: Easy 23/23, Medium 29/31, Hard 23/46.

Concurrency guidance: C4 is the practical balance (238 tok/s, 48.22% acceptance); C8 is the maximum measured throughput (276 tok/s, 48.38%).

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q5-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q5-LynnStyle.gguf \
  --mmproj Q5-LynnStyle/mmproj-Qwen3.8-27B-Q6_K.gguf \
  --model-draft Q5-LynnStyle/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q4-LynnStyle

Formal full-suite evaluation: GPQA 166/198 (83.84%) · MMLU 443/500 (88.60%) · LCB 74/100 (74.00%). LCB: Easy 23/23, Medium 30/31, Hard 21/46; 100 unique tasks, errors=0, clean gate PASS.

Concurrency guidance: C4 is the practical balance (257 tok/s, 47.69% acceptance); C8 is the maximum measured throughput (264 tok/s, 42.88%). The full sweep had zero timeout, request error, empty output, or punctuation collapse.

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q4-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q4-LynnStyle.gguf \
  --mmproj Q4-LynnStyle/mmproj-Qwen3.8-27B-Q6_K.gguf \
  --model-draft Q4-LynnStyle/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q3-LynnStyle

Formal full-suite evaluation: GPQA 172/198 (86.87%) · MMLU 435/500 (87.00%) · LCB 78/100 (78.00%). GPQA used 198 unique items with errors=0 and strict loop=0; MMLU had errors=0 and parse fail=0; LCB: Easy 23/23, Medium 28/31, Hard 27/46, with 100 unique tasks and errors=0. Every non-passing sample remains in the denominator.

Concurrency guidance: C4 is the practical balance (269 tok/s, 58.00% acceptance); C8 is the maximum measured throughput (286 tok/s, 52.41%). The full sweep had zero timeout, request error, empty output, or punctuation collapse.

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q3-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q3-LynnStyle.gguf \
  --mmproj Q3-LynnStyle/mmproj-Qwen3.8-27B-Q6_K.gguf \
  --model-draft Q3-LynnStyle/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q2-LynnStyle

Formal full-suite scores: GPQA 167/198 (84.34%) · MMLU 416/500 (83.20%) · LCB 75/100 (75.00%). LCB: Easy 23/23, Medium 29/31, Hard 23/46; all 23 length-limited 32K outputs remain failures in the denominator.

The final LCB view combines the byte-identical 99 unaffected rows from the original run with one Lynn-authorized exact retry of atcoder:arc181_b:81. The original request ended in a streaming JSON error. The retry used the same model, runner, request, and parameters and received a complete DONE, but again reached 32K with no submitted code, so the score remains 75/100. This is not represented as one uninterrupted error-free run.

Q2-LynnStyle uses GSQ-RCO mixed-precision quantization with IQ numerical refinement over 70 selected MLP matrices; attention, head, and SSM tensors were not numerically refined. If memory permits, Q3-LynnStyle or a higher tier remains the safer quality choice. This exact 12,999,977,600-byte build has no frozen TPS result, so throughput from the retired engineering Q2 is not reused.

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q2-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf \
  --mmproj Q2-LynnStyle/mmproj-Qwen3.8-27B-Q4_K_M.gguf \
  --model-draft Q2-LynnStyle/dflash2-qwen38-27b-Q4_K_M.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 4 --ctx-size 131072 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

The Q8_0, Q6_K, Q5-LynnStyle, Q4-LynnStyle, Q3-LynnStyle, and Q2-LynnStyle main GGUF files are GGUF v3, qwen35, 64-layer, 851-tensor text trunks with no embedded vision or MTP tensors; all passed remote byte/SHA256 readback. The published Q8_0 DFlash2 passed static parsing and a representative Q2-LynnStyle runtime smoke. Full per-tier concurrency retests remain pending, and historical DFlash2 grids are not attributed to this Q8_0 file.

LynnStyle mixed-precision quantization

Q5/Q4/Q3 LynnStyle tiers are quantized directly from the final SimPO F16 text trunk with the frozen 512-chunk imatrix (496 entries), without requantization. Protection combines T10 REAP, first/last-layer and SSM guards, plus per-tensor full-attention, SSM, and FFN audits; higher precision wins when rules overlap.

• Q5-LynnStyle (published): Q5 base; Q8 for L0–3, L60–63, and critical REAP layers 50/52/54/56/58; Q6 for high-sensitivity layers 21/42/44/46/48/49/53/57/59, full-attention q/k/v/o, and selected middle ssm_out. Of 498 converted tensors: Q8_0 104, Q6_K 186, Q5_K 208.
• Q4-LynnStyle (published): Q4 base; Q8 for L0–3 and L60–63; Q6 for critical shoulder layers 50/52/54/56–59; Q5 for high layers 21/42/44/46/48/49/53; Q6 for full-attention q/k/v/o and selected ssm_out. Distribution: Q8 62, Q6 156, Q5 172, Q4 108.
• Q3-LynnStyle (published): Q3 base; Q6 for L0–3 and L60–63; Q5 for L4–6 and 50/52/54/56–59; Q4 for high layers 21/42/44/46/48/49/53; Q5 for full-attention and middle ssm_out. Distribution: Q8 2, Q6 62, Q5 157, Q4 199, Q3 78.

Hub sidebars may expose only nominal GGUF general.file_type, which cannot fully describe custom per-tensor mixed precision. Treat each tier's manifest.json, SHA256SUMS, and actual GGUF parse as authoritative.

Q6_K

Formal full-suite evaluation: GPQA 171/198 (86.36%) · MMLU 440/500 (88.00%) · LCB 78/100 (78.00%).

LCB difficulty split: Easy 23/23, Medium 30/31, Hard 25/46.

Request errors, timeouts, empty outputs, and punctuation collapse were zero at all four tested concurrency levels. This serving sweep characterizes throughput and speculative acceptance; it does not replace formal quality evaluation.

Files

Recommended launch

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q6_K/Qwen3.8-27B-EfficientThink-SimPO-Q6_K.gguf \
  --mmproj Q6_K/mmproj-Qwen3.8-27B-Q8_0.gguf \
  --model-draft Q6_K/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

中文说明

已审查的 Q2–Q8 测评中未发现严格死循环。

BF16 / FP8 主仓：nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2

量化档位实测总览

能力分数只列正式全量冻结结果，所有未通过样本均保留在分母。Q8_0、Q6_K、Q5-LynnStyle、Q4-LynnStyle、Q3-LynnStyle 与 Q2-LynnStyle 均已发布，并完成四仓远端路径、字节和 LFS SHA256 回读。

Lynn Agent v0.87.0

Lynn Agent v0.87.0 已采用本系列 Q2-LynnStyle / Q3-LynnStyle + DFlash2。该组合已在 DGX Spark 实测通过；Mac Apple Silicon/Intel 公证、Windows 安装包运行检查、两仓 CI、三仓 main/tag 一致性、23 个公网文件完整 SHA256 与远程 CLI 安装均已通过。

自然摇曳的枝叶投影与柔和窗光，默认开启；悬停顶部‘树影’查看关闭路径，点击直达设置。后台暂停，减少动态效果时静止。图片已并入文件筛选，斜杠模板取代常驻任务模式，翻译移入消息菜单，专家圆桌改为可选插件，并修复会话编辑目标与停止预处理。Kimi Datasource 继续保留在 MCP 中，用户需自行扫码登录自己的账号。

本轮客户端更新未改变本仓模型权重、量化文件、测评分数或性能指标。

发布记录：GitHub 主仓 · GitHub 旧仓 · Gitee · CLI 包

llama.cpp 版本与启动设置

使用包含 2026-08-27 合入的官方 DFlash2 支持 的 llama.cpp：提交 b10f9ca58c89 或兼容的新版本。以下从官方仓构建最低支持版本：

git clone https://github.com/ggml-org/llama.cpp.git llama.cpp-efficientthink
git -C llama.cpp-efficientthink checkout b10f9ca58c89ccfc3653ac01e979dd085d582b76
cmake -S llama.cpp-efficientthink -B llama.cpp-efficientthink/build -DGGML_CUDA=ON
cmake --build llama.cpp-efficientthink/build --config Release -j 4 --target llama-server
export PATH="$PWD/llama.cpp-efficientthink/build/bin:$PATH"
llama-server --version
llama-server --help | grep -- draft-dflash

下列分档命令在本仓下载根目录执行；使用主仓镜像时先 cd GGUF。一次只选一个档位、启动一个服务。示例明确选用 GPU 0；只有自行配置多卡时才移除或调整 CUDA_VISIBLE_DEVICES=0。需安装支持所用 GPU 的 CUDA Toolkit 与编译器；GGUF 不依赖 Spark FP8 的 Docker 镜像。

实测布局为 8 slots × 32,768 tokens，总上下文 262144（--ctx-size 262144），不是八槽共用总计 32K。C4/C8 表示对八槽服务发出四/八个并发请求。显存较小时应同时缩小槽数和总上下文，例如单槽 --parallel 1 --ctx-size 32768；这种配置不能沿用卡中 C4/C8 的吞吐结果。权重文件大小不等于运行所需全部内存。

DFlash2 可选择独立 Q8_0 或 Q4_K_M draft、--spec-type draft-dflash、--spec-draft-n-max 7 与 --spec-draft-n-min 0，不是模型内置 MTP 路线。--jinja 和显式 XH 模板参数保留思考模式。默认只监听 127.0.0.1:8080，开放其他接口前请设置访问控制。--mmproj 用于图像；纯文本服务可去掉该行。

官方参考：服务参数、投机解码、CUDA 构建。修正后的命令对应官方参数与已测槽位布局；构建说明不代表重新跑过六档能力测评。

为什么选择 DFlash2，以及随包提供的 MTP 方案

本仓仍优先采用 DFlash2，主要看重实测草稿接受度和输出 TPS。 并行生成候选块、每轮验证接受多个 token，有助于减少逐 token 顺序解码开销。方法说明：DFlash 作者。

例如，Q8_0 的 C4 实测达到 60.98% 接受率、246 聚合 tok/s；Q3-LynnStyle 的 C4 达到 58.00%、269 聚合 tok/s。下面按各量化档分别列出数据：C4 为实用平衡档，C8 为最大实测吞吐档。接受度与 TPS 取决于模型、硬件、请求和并发；这些数据不是与 MTP 的同条件对照。

每档已随包提供 Q4_0 与 Q8_0 MTP GGUF

Q2-LynnStyle 至 Q8_0 的每个目录都包含同一组经过独立验证的 MTP sidecar：

两文件来自 ggml-org/Qwen3.8-27B-GGUF@0669b98607d47046c7c2b3f801011d54a08cfccf，均解析为 65 blocks、18 tensors 的 Qwen MTP GGUF。DGX Spark 已将两者分别与本仓 Q3-LynnStyle 主模型做真实兼容性 smoke：均成功加载、HTTP 200、生成非空最终内容并正常停止。该结果只证明兼容性，不是冻结的 MTP TPS 或能力成绩。

每次启动只选一条投机解码路线，不要同时启用 DFlash2 和 MTP。更小的 Q4_0 MTP 示例：

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q3-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q3-LynnStyle.gguf \
  --model-draft Q3-LynnStyle/mtp-Qwen3.8-27B-Q4_0.gguf \
  --spec-type draft-mtp \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 1 --ctx-size 32768 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

如需 Q8_0 MTP，将 draft 文件替换为 mtp-Qwen3.8-27B-Q8_0.gguf。只有图像服务才需要额外加入该档的 --mmproj。

分档 DFlash2 并发性能

统一实测口径：单卡主模型、DFlash2、--parallel 8 --ctx-size 262144（8 slots × 32K），C1/C2/C4/C8，256-token 定长压力输出，每请求最多 60 秒；DFlash2 使用 n_max=7、n_min=0。吞吐均为聚合 completion tok/s。

Q8_0

• **并发建议：**C4 实用平衡；C8 最大吞吐。
• 每档 15 个定长请求；超时、请求错误、空输出、标点坍塌均为 0。

Q6_K

• **并发建议：**C4 低延迟平衡；C8 最大吞吐。
• 整轮扫参中，超时、请求错误、空输出、标点坍塌均为 0。

Q5-LynnStyle

• **并发建议：**C4 实用平衡；C8 最大吞吐。
• 整轮扫参中，超时、请求错误、空输出、标点坍塌均为 0。

Q4-LynnStyle

• **并发建议：**C4 实用平衡；C8 最大吞吐。
• 整轮扫参中，超时、请求错误、空输出、标点坍塌均为 0。

Q3-LynnStyle

• **并发建议：**C4 为实用平衡档；C8 为最大实测吞吐。
• 整轮扫参中，超时、请求错误、空输出、标点坍塌均为 0。

训练方法

Qwen/Qwen3.8-27B → 能力保持 SFT → 终止行为 SimPO → 逐 tensor FP32 delta 合并 → BF16 → GGUF 量化。

• **SFT：**1,905 条样本，1 epoch，239 optimizer steps，effective batch 8；LoRA r=16、alpha=32、dropout=0.05；LR 5e-6，warmup 12 steps，seed 20260901。
• **SimPO：**110 组偏好对、73 个唯一 prompt，5 optimizer steps；beta=1.0、gamma=0.2、peak LR 5e-7；LoRA r=16、alpha=32、dropout=0，seed 20260903，world size 2，FSDP full sharding。
• **训练硬件：**双 NVIDIA RTX PRO 6000 Blackwell Server Edition。

本仓发布文本主 GGUF、视觉投影与经过独立验证的 DFlash2 draft。Q8_0 与 Q6_K 使用 Q8 视觉塔；Q5-LynnStyle、Q4-LynnStyle 与 Q3-LynnStyle 使用 Q6 视觉塔。六档都提供 Q8_0 与更小的 Q4_K_M DFlash2 draft，并另含 Q4_0 与 Q8_0 MTP sidecar。每次启动只选一条投机解码路线；这些 sidecar 都不是主模型。

已发布 GGUF 文件

Q8_0

正式全量评测：GPQA 164/198（82.83%）· MMLU 447/500（89.40%）· LCB 74/100（74.00%）。

**并发建议：**C4 为实用平衡档（246 tok/s，60.98% 接受率）；C8 为最大实测吞吐（256 tok/s，52.01%）。

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q8_0/Qwen3.8-27B-EfficientThink-SimPO-Q8_0.gguf \
  --mmproj Q8_0/mmproj-Qwen3.8-27B-Q8_0.gguf \
  --model-draft Q8_0/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q5-LynnStyle

正式全量评测：GPQA 164/198（82.83%）· MMLU 438/500（87.60%）· LCB 75/100（75.00%）。LCB：Easy 23/23、Medium 29/31、Hard 23/46。

**并发建议：**C4 为实用平衡档（238 tok/s，48.22% 接受率）；C8 为最大实测吞吐（276 tok/s，48.38%）。

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q5-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q5-LynnStyle.gguf \
  --mmproj Q5-LynnStyle/mmproj-Qwen3.8-27B-Q6_K.gguf \
  --model-draft Q5-LynnStyle/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q4-LynnStyle

正式全量评测：GPQA 166/198（83.84%）· MMLU 443/500（88.60%）· LCB 74/100（74.00%）。LCB：Easy 23/23、Medium 30/31、Hard 21/46；100 个唯一题目、errors=0、clean gate PASS。

**并发建议：**C4 为实用平衡档（257 tok/s，47.69% 接受率）；C8 为最大实测吞吐（264 tok/s，42.88%）。整轮扫参中，超时、请求错误、空输出、标点坍塌均为 0。

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q4-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q4-LynnStyle.gguf \
  --mmproj Q4-LynnStyle/mmproj-Qwen3.8-27B-Q6_K.gguf \
  --model-draft Q4-LynnStyle/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q3-LynnStyle

正式全量评测：GPQA 172/198（86.87%）· MMLU 435/500（87.00%）· LCB 78/100（78.00%）。GPQA 为 198 个唯一题目、errors=0、strict loop=0；MMLU errors=0、parse fail=0；LCB：Easy 23/23、Medium 28/31、Hard 27/46，100 个唯一题目、errors=0。所有未通过样本均保留在分母。

**并发建议：**C4 为实用平衡档（269 tok/s，58.00% 接受率）；C8 为最大实测吞吐（286 tok/s，52.41%）。整轮扫参中，超时、请求错误、空输出、标点坍塌均为 0。

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q3-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q3-LynnStyle.gguf \
  --mmproj Q3-LynnStyle/mmproj-Qwen3.8-27B-Q6_K.gguf \
  --model-draft Q3-LynnStyle/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q2-LynnStyle

正式全量成绩：GPQA 167/198（84.34%）· MMLU 416/500（83.20%）· LCB 75/100（75.00%）。LCB：Easy 23/23、Medium 29/31、Hard 23/46；23 个 32K 长度结束均以失败计入。

LCB 最终视图由原始运行中保持逐字节不变的 99 题，加上 Lynn 授权对 atcoder:arc181_b:81 的一次精确补测组成。该题原运行发生流式 JSON 错误；补测使用同一模型、runner、请求与参数，完整收到 DONE，但仍在 32K 结束且没有提交代码，因此总分仍为 75/100。该结果不是一次连续无故障运行。

Q2-LynnStyle 使用 GSQ-RCO 混合精度量化，并以 IQ 数值细化优化 70 个选定 MLP 矩阵；attention、head 与 SSM tensor 未做数值细化。若内存允许，Q3-LynnStyle 及更高档位仍是更稳妥的质量选择。此精确 12,999,977,600-byte 构建没有冻结 TPS，不能套用已淘汰工程 Q2 的吞吐。

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q2-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf \
  --mmproj Q2-LynnStyle/mmproj-Qwen3.8-27B-Q4_K_M.gguf \
  --model-draft Q2-LynnStyle/dflash2-qwen38-27b-Q4_K_M.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 4 --ctx-size 131072 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080

Q8_0、Q6_K、Q5-LynnStyle、Q4-LynnStyle、Q3-LynnStyle 与 Q2-LynnStyle 的主 GGUF 均为 GGUF v3、qwen35、64 层、851 tensors，主文件不内嵌视觉或 MTP tensor；均已通过远端字节/SHA256 回读。当前发布的 Q8_0 DFlash2 已通过静态解析与 Q2-LynnStyle 代表性运行 smoke；其余分档完整并发重测待补，旧 DFlash2 数据不归因于该 Q8_0 文件。

LynnStyle 混合精度量化

Q5/Q4/Q3 LynnStyle 均从最终 SimPO F16 文本主干直接量化，使用冻结的 512-chunk imatrix（496 entries），不做二次 requantize。保护策略综合 T10 REAP、首尾层/SSM 保护与 full-attention、SSM、FFN 的逐 tensor 审计；规则冲突时采用更高精度。

• **Q5-LynnStyle（已发布）：**基础 Q5；L0–3、L60–63 与关键 REAP 层 50/52/54/56/58 使用 Q8；高敏感层 21/42/44/46/48/49/53/57/59 使用 Q6；full-attention q/k/v/o 与选定中层 ssm_out 使用 Q6。498 个转换 tensor：Q8_0 104、Q6_K 186、Q5_K 208。
• **Q4-LynnStyle（已发布）：**基础 Q4；L0–3、L60–63 使用 Q8；关键肩部层 50/52/54/56–59 使用 Q6；高敏感层 21/42/44/46/48/49/53 使用 Q5；full-attention q/k/v/o 与选定 ssm_out 使用 Q6。分布：Q8 62、Q6 156、Q5 172、Q4 108。
• **Q3-LynnStyle（已发布）：**基础 Q3；L0–3、L60–63 使用 Q6；L4–6、50/52/54/56–59 使用 Q5；高敏感层 21/42/44/46/48/49/53 使用 Q4；full-attention 与中层 ssm_out 使用 Q5。分布：Q8 2、Q6 62、Q5 157、Q4 199、Q3 78。

Hub 侧边栏可能仅显示 GGUF 名义 general.file_type，无法完整表达自定义逐 tensor 混合精度；以各目录 manifest.json、SHA256SUMS 与实际 GGUF parse 为准。

Q6_K

正式全量评测：GPQA 171/198（86.36%）· MMLU 440/500（88.00%）· LCB 78/100（78.00%）。

LCB 难度分布：Easy 23/23、Medium 30/31、Hard 25/46。

四档固定输出测试的请求错误、超时、空输出和标点坍塌均为 0。吞吐测试不代替正式质量评测。

文件

推荐启动

CUDA_VISIBLE_DEVICES=0 llama-server \
  --model Q6_K/Qwen3.8-27B-EfficientThink-SimPO-Q6_K.gguf \
  --mmproj Q6_K/mmproj-Qwen3.8-27B-Q8_0.gguf \
  --model-draft Q6_K/dflash2-qwen38-27b-Q8_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --gpu-layers all --gpu-layers-draft all \
  --parallel 8 --ctx-size 262144 --flash-attn on \
  --jinja --chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --host 127.0.0.1 --port 8080