Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer

Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer

在 Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2(官方 Qwen3.8-27B 经 SFT + SimPO 后训练得到的模型)基础上继续做多轮 SFT 与 RLOO 后训练,专治原版「已有答案却停不下来、写满 94K 也不给最终答案」的问题。本仓库是 **GGUF** 档的独立仓库,目前提供 GGUF/ 下的八个 GGUF 包:**Q2 / Q3 / Q4 LynnStyle**(内置 **Q4 0** MTP 头)、同样三档的 **-Q8MTP** 版(内置 **Q8 0** MTP 头)、**Q6 K** 与 **Q8 0**(均内置 MTP 头),以及 GGUF-NInfer/ 下对应的 NInfer 包(已内置视觉)。另附 BF16 视觉塔原始文件( GGUF-NInfer/vision-bf16.safetensors ),供自行打包/转换;V3 .ninfer 已内置视觉,运行时不需要。llama.cpp 多模态请用 --mmproj GGUF/mmproj-Qwen3.8-27B-Q8 0.gguf (mmproj 只在 GGUF/ ,不要求放进 NInfer 包)。BF16、静态 FP8、NVFP4、NInfer W4A4、INT8 W8A8、INT4 W4A16 等其他档见主仓库 Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2。

• *对照基线**:原版 Qwen3.8 27B（FP8）三项为 GPQA 177/198、MMLU 444/500、LCB 83/100;GPQA 思考长度(P50 / P70 / P90 token)为 5,299 / 12,607 / 50,577。
• *GGUF Q2 LynnStyle · 成绩**:GPQA 177/198、MMLU 433/500、LCB 90/100(llama.cpp,C4,100K 全量;成绩在内置 MTP 前测得(外挂 Q8 MTP 草稿),量化权重相同)。
• *GGUF Q2 LynnStyle · 思考长度(P50 / P70 / P90 token)**:GPQA 2,989.5 / 8,923.8 / 27,156.1;MMLU 181 / 326.3 / 1,020.5;LCB 5,656.5 / 17,835.8 / 40,067.3。
• *GGUF Q2 LynnStyle · 速度**:NInfer 内置 Q4 MTP,C8 合计 521.0 tok/s(生成上限 512 短测);llama.cpp 内置 Q4 MTP,C4 合计 174.8 tok/s(生成上限 4096 长测)。口径不同,勿直接横比。
• *GGUF Q2 LynnStyle · 体积**:GGUF 13,276,009,792 字节(BPW 3.89),NInfer 13,568,617,472 字节,含视觉(文本 + MTP BPW 3.89)。
• *GGUF Q3 LynnStyle-16G · 简介**:我们最低的 Q3 档,推荐 16G 显存( GGUF/Qwen3.8-27B-Coder390-Q3LynnStyle16G-MTP.gguf )。相对此前的 Q2 方案做了调整:最低精度改为 IQ3 XXS(主体无 Q6 / Q8),并做了重要性矩阵校正和 GSQ-RCO(只采用同字节的 GSQ 精修,体积不变)。内置 Q4 K MTP 头(norm 为 F32)。GGUF 只含文本,看图用 GGUF/ 内的 mmproj。同权重 NInfer 包在 GGUF-NInfer/ ,内置 Q8 视觉塔 + Q4 K MTP。
• *GGUF Q3 LynnStyle-16G · 成绩**:GPQA 172/198、MMLU 427/500、LCB 90/100(llama.cpp + 外挂 DFlash2 草稿,C4,100K 上下文,生成上限 94,208,不计超时,reasoning effort xhigh,全量)。
• *GGUF Q3 LynnStyle-16G · 思考长度(P50 / P70 / P90 token)**:GPQA 3,811 / 11,404 / 31,418;MMLU 190 / 361 / 1,665;LCB 6,520 / 21,586 / 43,184。GPQA 用 reasoning tokens ;MMLU 和 LCB 的记录没有单独的 reasoning 字段,用的是 completion tokens (含最终答案)。
• *GGUF Q3 LynnStyle-16G · 体积**:GGUF 13,276,009,920 字节(约 12.4 GiB,BPW 3.89);NInfer 13,765,136,896 字节(含 Q8 视觉,约 12.8 GiB,文本 + MTP BPW 3.89)。
• *GGUF Q3 LynnStyle-16G · 速度**:在一张 RTX PRO 6000 Blackwell 上测得。llama.cpp 融合 MTP,生成上限 4096(长测):C1 31.9 / C2 46.6 / C4 78.1 tok/s,接受率 0.597。NInfer,生成上限 512(短测):C1 33.5 / C2 44.7 / C4 77.4 / C8 127.0 tok/s,接受率 0.488。长短测协议不同,不要直接横比。
• *GGUF Q3 LynnStyle · 成绩**:GPQA 173/198、MMLU 449/500、LCB 92/100(llama.cpp,C4,100K 全量;成绩在内置 MTP 前测得(外挂 Q8 MTP 草稿),量化权重相同)。
• *GGUF Q3 LynnStyle · 思考长度(P50 / P70 / P90 token)**:GPQA 3,125.5 / 7,265.2 / 24,343.1;MMLU 148 / 279.3 / 923.4;LCB 4,421 / 14,232.2 / 34,946。
• *GGUF Q3 LynnStyle · 速度**:NInfer 内置 Q4 MTP,C8 合计 512.3 tok/s(生成上限 512 短测);llama.cpp 内置 Q4 MTP,C4 合计 191.8 tok/s(生成上限 4096 长测)。口径不同,勿直接横比。
• *GGUF Q3 LynnStyle · 体积**:GGUF 17,303,770,432 字节(BPW 5.07),NInfer 17,596,369,920 字节,含视觉(文本 + MTP BPW 5.07)。
• *GGUF Q4 LynnStyle · 成绩**:GPQA 175/198、MMLU 448/500、LCB 89/100(llama.cpp,C4,100K 全量;成绩在内置 MTP 前测得(外挂 Q8 MTP 草稿),量化权重相同)。
• *GGUF Q4 LynnStyle · 思考长度(P50 / P70 / P90 token)**:GPQA 2,951.0 / 8,159.3 / 25,649.0;MMLU 164 / 312.9 / 1,077.7;LCB 3,669.5 / 14,810.5 / 38,182.1。
• *GGUF Q4 LynnStyle · 速度**:NInfer 内置 Q4 MTP,C8 合计 436.4 tok/s(生成上限 512 短测);llama.cpp 内置 Q4 MTP,C4 合计 81.1 tok/s(生成上限 4096 长测)。口径不同,勿直接横比。
• *GGUF Q4 LynnStyle · 体积**:GGUF 19,620,406,592 字节(BPW 5.75),NInfer 19,913,006,080 字节,含视觉(文本 + MTP BPW 5.74)。
• *GGUF Q6 K · 成绩**:GPQA 174/198、MMLU 448/500、LCB 91/100(llama.cpp,C4,100K 全量)。
• *GGUF Q6 K · 思考长度(P50 / P70 / P90 token)**:GPQA 3,683 / 9,814 / 27,966。
• *GGUF Q6 K · 速度**:NInfer 开 MTP,C4 合计 235.5 tok/s;llama.cpp 开 MTP,C8 合计 189.0 tok/s(均为生成上限 512 快测)。
• *GGUF Q6 K · 体积**:GGUF 23,177,516,384 字节(BPW 6.79),NInfer 23,379,962,880 字节,含视觉(文本 + MTP BPW 6.76)。
• *GGUF Q8 0 · 成绩**:GPQA 171/198、MMLU 443/500、LCB 90/100(llama.cpp,C4,100K 全量)。
• *GGUF Q8 0 · 思考长度(P50 / P70 / P90 token)**:GPQA 3,102.5 / 7,971 / 28,891.6;MMLU 164 / 291 / 741;LCB 4,098.5 / 16,087.8 / 39,135.5。
• *GGUF Q8 0 · 速度**:NInfer 开 MTP,C8 合计 486.3 tok/s(同口径测速);llama.cpp 开 MTP,C4 合计 190.8 tok/s。
• *GGUF Q8 0 · 体积**:GGUF 29,069,202,688 字节(BPW 8.51),NInfer 29,361,802,240 字节,含视觉(文本 + MTP BPW 8.51)。

BPW 分母均为文本 + MTP 参数数 27,320,697,856。

相关仓库**

主仓（BF16 / FP8 / NVFP4 / INT8 / INT4 / GGUF 等全部版本）：Hugging Face · ModelScope

NVFP4-NInfer（NVFP4 与 NInfer W4A4 / W4A4-W8A8，内置 MTP 与 DFlash2；另有 INT8 W8A8 与 INT4 W4A16）：Hugging Face · ModelScope

训练方法

谱系**:原版 Qwen3.8 27B(FP8;177 / 444 / 83)→ 我们的 SFT + SimPO110 终版,即 EfficientThink 基座(171 / 442 / 89)→ K3 续训 SFT → 第二周 SFT(week2dose,update-225)→ 第一轮 RLOO(182 组)→ 再一轮 SFT(merge-sft-100),得 sft-base-rloo(177 / 448 / 90)→ 第二轮 RLOO(172 组)→ **Coder390**(178 / 445 / 90,动态 FP8 口径)。括号内依次为 GPQA / MMLU / LCB,均为 100K 同口径全量。

训练基座**:本模型以我们自己的 EfficientThink SFT + SimPO 模型(SimPO110 终版)为起点,继续做多轮 SFT 与 RLOO,该基座模型本身是官方 Qwen3.8-27B 的后训练版本,见 Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2。

名字片段 含义

• -- ---

Coder 代码强化

390 GPQA、MMLU、LCB 三项都到 90 分

EfficientThink 训练目标:去掉无效的长推理尾巴,保留必要的长推理

Opus5.5、GPT6Astra 撰写出题金标与教师模型轨迹

Grok4.7 本次训练的主持人,全程核对、筛选数据

DSV4Pro、K3 其轨迹作为教师模型用于训练;K3 还做了 RLOO 价值评审和一部分出题工作

SFT-RLOO 训练方法:SFT(含 SimPO)基座上,SFT 与 RLOO 交替进行

GGUF-NInfer 本仓库内容:GGUF Q2 / Q3 / Q4 LynnStyle(含 Q8 MTP 版)、Q6 K、Q8 0 包(均内置 MTP 头),以及由它们转换的 NInfer 包(已内置视觉)

要解决的问题**:模型在局部已经有答案之后,仍用 Wait / Actually 把同一套推导反复再推,直到写满上下文,最终答案为空。这类题多数不是不会,而是停不下来。

RLOO 数据**:每题采样 8 条轨迹,只保留组内有对有错的题(8 条全对的不入训);组内 0 或 1 条对的,补 1 条审过的短教师轨迹,替换该组最短的一条错轨迹(共 27 组)。最终 172 组、1,376 条:正确 859 条、错误 517 条,其中空答 68 条。

奖励**:惩罚只打在错的一侧,长而正确的推理仍得正奖励,以免压制必要的长推理。

情形 奖励

• -- ---:

短且对(

Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer

A model post-trained with several rounds of SFT and RLOO on top of Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2, which is itself the official Qwen3.8-27B post-trained with SFT and SimPO. It is built to fix the original model's habit of failing to stop: it already has an answer, yet keeps re-deriving until it hits the 94K cap with no final answer. This repository is the standalone **GGUF** repository with eight GGUF packages under GGUF/ : **Q2 / Q3 / Q4 LynnStyle** (built-in **Q4 0** MTP head), the same three as **-Q8MTP** versions (built-in **Q8 0** MTP head), **Q6 K**, and **Q8 0** (each with a built-in MTP head), plus the matching NInfer packages under GGUF-NInfer/ (vision built in). The raw BF16 vision-tower file ( GGUF-NInfer/vision-bf16.safetensors ) is included for your own packaging/conversion; V3 .ninfer packages already have vision built in and do not need it at runtime. For llama.cpp multimodal, use --mmproj GGUF/mmproj-Qwen3.8-27B-Q8 0.gguf (Q8 0 mmproj in GGUF/ only; not required inside NInfer packages). The BF16, static FP8, NVFP4, NInfer W4A4, INT8 W8A8, INT4 W4A16, and other tiers are in the main repository Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2.

• *Baseline**: the original Qwen3.8 27B (FP8) scores GPQA 177/198, MMLU 444/500, LCB 83/100; its GPQA reasoning length (P50 / P70 / P90 tokens) is 5,299 / 12,607 / 50,577.
• *GGUF Q2 LynnStyle · scores**: GPQA 177/198, MMLU 433/500, LCB 90/100 (llama.cpp, C4, 100K full sets; scored before the built-in MTP head was attached (external Q8 MTP draft); same quantized weights).
• *GGUF Q2 LynnStyle · reasoning length (P50 / P70 / P90 tokens)**: GPQA 2,989.5 / 8,923.8 / 27,156.1; MMLU 181 / 326.3 / 1,020.5; LCB 5,656.5 / 17,835.8 / 40,067.3.
• *GGUF Q2 LynnStyle · speed**: NInfer with built-in Q4 MTP, C8 aggregate 521.0 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 174.8 tok/s (4096-token cap, long run). Protocols differ; do not compare directly.
• *GGUF Q2 LynnStyle · size**: GGUF 13,276,009,792 bytes (BPW 3.89), NInfer 13,568,617,472 bytes with vision (text + MTP BPW 3.89).
• *GGUF Q3 LynnStyle-16G · what it is**: our lowest Q3 tier, recommended for 16 GB VRAM ( GGUF/Qwen3.8-27B-Coder390-Q3LynnStyle16G-MTP.gguf ). It changes the earlier Q2 scheme: the lowest type is now IQ3 XXS (no Q6 / Q8 in the body), built with importance-matrix correction and GSQ-RCO (only same-byte GSQ refinements were applied, so the size is unchanged). Built-in Q4 K MTP head (norms F32). Text-only GGUF; vision via the mmproj in GGUF/ . Matching NInfer package under GGUF-NInfer/ with built-in Q8 vision tower + Q4 K MTP.
• *GGUF Q3 LynnStyle-16G · scores**: GPQA 172/198, MMLU 427/500, LCB 90/100 (llama.cpp + DFlash2 external draft, C4, 100K context, max tokens 94,208, no timeout, reasoning effort xhigh; full sets).
• *GGUF Q3 LynnStyle-16G · reasoning length (P50 / P70 / P90 tokens)**: GPQA 3,811 / 11,404 / 31,418; MMLU 190 / 361 / 1,665; LCB 6,520 / 21,586 / 43,184. GPQA uses reasoning tokens ; the MMLU and LCB records have no separate reasoning field, so completion tokens (including the final answer) is used.
• *GGUF Q3 LynnStyle-16G · size**: GGUF 13,276,009,920 bytes (~12.4 GiB, BPW 3.89); NInfer 13,765,136,896 bytes with Q8 vision (~12.8 GiB, text + MTP BPW 3.89).
• *GGUF Q3 LynnStyle-16G · speed**: measured on one RTX PRO 6000 Blackwell. llama.cpp fused MTP, 4096-token generation cap (long): C1 31.9 / C2 46.6 / C4 78.1 tok/s, acceptance 0.597. NInfer, 512-token generation cap (short): C1 33.5 / C2 44.7 / C4 77.4 / C8 127.0 tok/s, acceptance 0.488. Protocols differ (long vs short); do not compare directly.
• *GGUF Q3 LynnStyle · scores**: GPQA 173/198, MMLU 449/500, LCB 92/100 (llama.cpp, C4, 100K full sets; scored before the built-in MTP head was attached (external Q8 MTP draft); same quantized weights).
• *GGUF Q3 LynnStyle · reasoning length (P50 / P70 / P90 tokens)**: GPQA 3,125.5 / 7,265.2 / 24,343.1; MMLU 148 / 279.3 / 923.4; LCB 4,421 / 14,232.2 / 34,946.
• *GGUF Q3 LynnStyle · speed**: NInfer with built-in Q4 MTP, C8 aggregate 512.3 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 191.8 tok/s (4096-token cap, long run). Protocols differ; do not compare directly.
• *GGUF Q3 LynnStyle · size**: GGUF 17,303,770,432 bytes (BPW 5.07), NInfer 17,596,369,920 bytes with vision (text + MTP BPW 5.07).
• *GGUF Q4 LynnStyle · scores**: GPQA 175/198, MMLU 448/500, LCB 89/100 (llama.cpp, C4, 100K full sets; scored before the built-in MTP head was attached (external Q8 MTP draft); same quantized weights).
• *GGUF Q4 LynnStyle · reasoning length (P50 / P70 / P90 tokens)**: GPQA 2,951.0 / 8,159.3 / 25,649.0; MMLU 164 / 312.9 / 1,077.7; LCB 3,669.5 / 14,810.5 / 38,182.1.
• *GGUF Q4 LynnStyle · speed**: NInfer with built-in Q4 MTP, C8 aggregate 436.4 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 81.1 tok/s (4096-token cap, long run). Protocols differ; do not compare directly.
• *GGUF Q4 LynnStyle · size**: GGUF 19,620,406,592 bytes (BPW 5.75), NInfer 19,913,006,080 bytes with vision (text + MTP BPW 5.74).
• *GGUF Q6 K · scores**: GPQA 174/198, MMLU 448/500, LCB 91/100 (llama.cpp, C4, 100K full sets).
• *GGUF Q6 K · reasoning length (P50 / P70 / P90 tokens)**: GPQA 3,683 / 9,814 / 27,966.
• *GGUF Q6 K · speed**: NInfer with MTP, C4 aggregate 235.5 tok/s; llama.cpp with MTP, C8 aggregate 189.0 tok/s (both 512-token-cap quick runs).
• *GGUF Q6 K · size**: GGUF 23,177,516,384 bytes (BPW 6.79), NInfer 23,379,962,880 bytes with vision (text + MTP BPW 6.76).
• *GGUF Q8 0 · scores**: GPQA 171/198, MMLU 443/500, LCB 90/100 (llama.cpp, C4, 100K full sets).
• *GGUF Q8 0 · reasoning length (P50 / P70 / P90 tokens)**: GPQA 3,102.5 / 7,971 / 28,891.6; MMLU 164 / 291 / 741; LCB 4,098.5 / 16,087.8 / 39,135.5.
• *GGUF Q8 0 · speed**: NInfer with MTP, C8 aggregate 486.3 tok/s (same-caliber bench); llama.cpp with MTP, C4 aggregate 190.8 tok/s.
• *GGUF Q8 0 · size**: GGUF 29,069,202,688 bytes (BPW 8.51), NInfer 29,361,802,240 bytes with vision (text + MTP BPW 8.51).

The BPW denominator is always the text + MTP parameter count 27,320,697,856.

Related repositories**

Main repository (BF16 / FP8 / NVFP4 / INT8 / INT4 / GGUF and all other tiers): Hugging Face · ModelScope

NVFP4-NInfer (NVFP4 and NInfer W4A4 / W4A4-W8A8 with built-in MTP and DFlash2, plus INT8 W8A8 and INT4 W4A16): Hugging Face · ModelScope

Training method

Lineage**: original Qwen3.8 27B (FP8; 177 / 444 / 83) → our SFT + SimPO110 final, i.e. the EfficientThink base (171 / 442 / 89) → K3 continuation SFT → week-2 SFT (week2dose, update-225) → first RLOO round (182 groups) → another SFT round (merge-sft-100), giving sft-base-rloo (177 / 448 / 90) → second RLOO round (172 groups) → **Coder390** (178 / 445 / 90, dynamic FP8). Numbers are GPQA / MMLU / LCB, all full suites under the same 100K protocol.

Training base**: this model starts from our own EfficientThink SFT + SimPO model (the SimPO110 final) and goes through several rounds of SFT and RLOO, which is itself a post-trained version of the official Qwen3.8-27B: Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2.

Name part Meaning

• -- ---

Coder Strengthened coding

390 GPQA, MMLU, and LCB all reach 90

EfficientThink Training goal: remove unproductive reasoning tails while keeping necessary long reasoning

Opus5.5, GPT6Astra Wrote the gold answers and teacher trajectories

Grok4.7 Host of this training run; checked and filtered data throughout

DSV4Pro, K3 Their trajectories served as teacher-model data; K3 also performed the RLOO value review and wrote part of the problems

SFT-RLOO Method: alternating SFT and RLOO rounds on an SFT (incl. SimPO) base

GGUF-NInfer Contents of this repository: GGUF Q2 / Q3 / Q4 LynnStyle (plus Q8 MTP versions), Q6 K, and Q8 0 packages, each with a built-in MTP head, and the NInfer packages converted from them (vision built in)

The problem**: after the model already has a local answer, it keeps re-running the same derivation with Wait / Actually until the context is full, and the final answer is empty. In most of these cases the model can solve the problem; it just does not stop.

RLOO data**: 8 trajectories sampled per problem; only groups with both correct and wrong trajectories are kept (all-correct groups are dropped). Groups with 0 or 1 correct trajectory receive one reviewed short teacher trajectory, which replaces the shortest wrong trajectory in that group (27 groups). Final set: 172 groups, 1,376 trajectories: 859 correct and 517 wrong, including 68 empty answers.

Reward**: penalties apply only to wrong trajectories; long correct reasoning still earns a positive reward, so necessary long reasoning is not suppressed.

Case Reward

• -- ---:

Short and correct (