Recommended reasoning effort: use medium by default. For simple tasks, or if the model becomes repetitive or keeps writing after reaching a final answer, switch to low. Reserve xhigh for genuinely complex tasks and allow a sufficient reasoning/token budget.

Why: the official Qwen3.8 template injects an extra xhigh instruction to “validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity.” The low branch instead asks for brief, focused reasoning and a direct conclusion; medium keeps thinking enabled without injecting either extra effort instruction. In local FP8 and Q8 troubleshooting, changing to low immediately stopped the observed loop; this is operational evidence, not a universal quality guarantee.

Overview

This repository contains the complete BF16 merge and the verified Q8_0 llama.cpp package for the C0/T10 directional refusal-dissolve derivative of official Qwen3.8-27B. The Q8_0 main model, official Q8 MTP draft, and official Q8 vision projector are placed at repository root.

Research disclaimer: This experimental release is provided solely to study the technical feasibility and behavioral effects of refusal-tendency dissolution. It is not a comprehensive safety conclusion, an endorsement of unrestricted use, or professional advice. Users are responsible for lawful and appropriate use and for independently verifying model outputs.

Base model and derivation

ModelScope base: Qwen/Qwen3.8-27B
Hugging Face base: Qwen/Qwen3.8-27B
Frozen source snapshot: ccd1ef6e1a997cde3a9d544bda75d54e7297bb90
Derivation: official Qwen3.8-27B → C0 search → frozen T10 policy → BF16 merge → llama.cpp Q8_0
T10 parameters: rank 3, alpha 3, dropout 0; targets attn.o_proj, mlp.down_proj, and compatible out_proj modules.

Release artifacts

Role	Root file	Bytes	SHA256
BF16	`model-00001-of-00012.safetensors` … `model-00012-of-00012.safetensors`	54,736,634,773 total	See `SHA256SUMS`
Q8_0 main	`Qwen3.8-27B-Uncensored-Heretic-T10-Q8_0.gguf`	28,595,763,584	`fe17bff3e43463764b734b945d92c53974143193e2431056a929e3d1b95b5b58`
Q8 MTP draft	`Q8-MTP-draft.gguf`	3,164,006,752	`e2c0e242a52a92d16d67ef9ca99f41c3028e1cf34b321ecff6adb5a9a3c6ffca`
Q8 vision projector	`mmproj-Qwen3.8-27B-Q8_0.gguf`	629,247,072	`acaa21ef79fbc451b635c8db29db65b33329aa01c15b968c975a9460cfc82b5c`

The main Q8_0 GGUF is qwen35, file type 7, 851 tensors. The MTP sidecar is qwen35, file type 7, 18 tensors. The vision projector is clip, file type 7, 334 tensors.

Frozen Q8_0 results

Suite	Final score
MMLU500	435/500 (87.0%)
GPQA198	169/198 (85.35%)
LCB100	76/100
Refusal140 hard refusal	1/140 (0.7%)
Refusal140 soft disclaimer	32/140 (22.9%)

BF16 Refusal140 hard refusal: 0/140 (0.0%). The Q8_0 hard-refusal result remains 1/140 (0.7%).

LCB difficulty: easy 23/23, medium 30/31, hard 23/46. The structured result package is available at evaluation/q8-results-v1.json.

llama.cpp startup

Bare Q8_0 with vision:

llama-server   -m Qwen3.8-27B-Uncensored-Heretic-T10-Q8_0.gguf   --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf   -ngl all -c 32768 --host 127.0.0.1 --port 8080 --no-webui

Q8_0 with MTP and vision:

llama-server   -m Qwen3.8-27B-Uncensored-Heretic-T10-Q8_0.gguf   --model-draft Q8-MTP-draft.gguf   --spec-type draft-mtp   --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf   -ngl all -c 32768 --host 127.0.0.1 --port 8080 --no-webui

Validation and limits

Bare Q8_0 load/generation: PASS (SMOKE_OK_38).
Q8 MTP + vision load, text generation, and semantic color vision: PASS.
Strict synthetic OCR did not reproduce the requested string exactly; OCR quality remains a known limitation.
Qwen3.8 enables thinking by default. Use the bundled chat template and a recent Qwen3.8-capable llama.cpp build.
SHA256SUMS and manifest.json define the exact release.

License

Apache License 2.0. Users must also follow the official Qwen3.8-27B terms and applicable law.

中文说明

思考强度建议：默认使用 medium。简单任务，或出现重复循环、已经给出最终答案仍继续输出时，切换到 low。xhigh 仅用于确实复杂的任务，并预留足够的思考/token 预算。

**原因：**Qwen3.8 官方模板会在 xhigh 分支额外注入“验证关键假设，考虑合理且可能成立的替代方案，并优先确保正确性、一致性和表达清晰度”；low 分支要求保持思考简短、聚焦并直接得出结论；medium 保留 thinking，但不注入上述两类额外 effort 指令。本地 FP8 与 Q8 排障中，改为 low 后观察到的循环立即停止；这是运行证据，不代表所有任务的通用质量结论。

本仓库包含官方 Qwen3.8-27B 经 C0/T10 定向拒答溶解后的完整 BF16 合并权重，以及已验证的 Q8_0 llama.cpp 套件。主 Q8_0、官方 Q8 MTP draft 与官方 Q8 视觉塔均放在仓库根目录。

研究免责声明： 本实验版本仅用于研究拒答倾向溶解的技术可行性与行为影响。它不构成完整的安全结论，不代表对无限制使用的认可，也不构成任何专业建议。使用者有责任依法、适当地使用模型，并独立核验模型输出。

基座与派生关系

ModelScope 基座：Qwen/Qwen3.8-27B
Hugging Face 基座：Qwen/Qwen3.8-27B
冻结源快照：ccd1ef6e1a997cde3a9d544bda75d54e7297bb90
派生链：官方 Qwen3.8-27B → C0 搜索 → T10 冻结方案 → BF16 merge → llama.cpp Q8_0
T10 参数：rank 3、alpha 3、dropout 0；目标为 attn.o_proj、mlp.down_proj 和兼容的 out_proj。

根目录发布文件

角色	文件	Bytes	SHA256
BF16	`model-00001-of-00012.safetensors` … `model-00012-of-00012.safetensors`	54,736,634,773 总计	见 `SHA256SUMS`
Q8_0 主模型	`Qwen3.8-27B-Uncensored-Heretic-T10-Q8_0.gguf`	28,595,763,584	`fe17bff3e43463764b734b945d92c53974143193e2431056a929e3d1b95b5b58`
Q8 MTP draft	`Q8-MTP-draft.gguf`	3,164,006,752	`e2c0e242a52a92d16d67ef9ca99f41c3028e1cf34b321ecff6adb5a9a3c6ffca`
Q8 视觉塔	`mmproj-Qwen3.8-27B-Q8_0.gguf`	629,247,072	`acaa21ef79fbc451b635c8db29db65b33329aa01c15b968c975a9460cfc82b5c`

Q8_0 冻结成绩

测评	最终得分
MMLU500	435/500（87.0%）
GPQA198	169/198（85.35%）
LCB100	76/100
Refusal140 硬拒答	1/140（0.7%）
Refusal140 边答边免责	32/140（22.9%）

BF16 Refusal140 硬拒答：0/140（0.0%）。 Q8_0 硬拒答结果仍为 1/140（0.7%）。

LCB 分项：easy 23/23、medium 30/31、hard 23/46。结构化成绩见 evaluation/q8-results-v1.json。

llama.cpp 启动

上面的两条命令分别对应 bare Q8_0 + vision 与 Q8_0 + MTP + vision。MTP 命令必须同时保留 --model-draft Q8-MTP-draft.gguf 与 --spec-type draft-mtp。

验证与限制

bare Q8_0 真实 load/generation PASS。
Q8 MTP + 视觉塔真实 load、文本生成与颜色语义视觉 PASS。
严格合成 OCR 未准确复现目标字符串，OCR 输出质量仍有限制。
Qwen3.8 默认启用 thinking；请使用仓内 chat template 与支持 Qwen3.8 的较新 llama.cpp。
SHA256SUMS 与 manifest.json 定义本次发布的精确文件。

许可证

Apache License 2.0。使用时还需遵守官方 Qwen3.8-27B 许可与适用法律。

nerkyor/Qwen3.8-27B-Uncensored-Heretic-T10-BF16

Qwen3.8-27B Uncensored Heretic T10 BF16

English