@no_stp_on_snek: 经过测试该模型的一组提示词回放，我可以自信地说，它在对话角色扮演、指令遵循、线程讨论和交叉引用方面，完全可以替代我运行gpt-mini-5.1的某个产品。

测试了来自此模型的一组提示词回放后，我可以自信地说，它在会话角色扮演、指令遵循、线程化讨论及交叉引用方面，可以很好地替代我当前运行gpt-mini-5.1的一个产品。干得好！—— # llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF · Hugging Face 来源：https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF ## 🚨⚠️ 我已达到Hugging Face的免费存储上限 ⚠️🚨 除非我能承担额外存储的费用，否则无法再上传新模型。我作为独立贡献者托管了70+个免费模型，这项工作没有报酬。没有您的支持，将无法上传更多新模型。 🎉 Patreon（月付）（https://patreon.com/LLMfan46） | ☕ Ko-fi（一次性）（https://ko-fi.com/llmfan46） 每一笔捐赠都将直接用于Hugging Face存储费用，以保持模型对所有人免费。 — 这是完整模型，保留了全部20个MTP。 ### https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#88-fewer-refusals-10100-uncensored-vs-83100-original-while-preserving-model-quality-00015-kl-divergence 拒绝率降低88%（去审查版10/100 vs 原版83/100），同时保持了模型质量（KL散度0.0015）。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#%E2%9D%A4%EF%B8%8F-support-my-work ❤️ 支持我的工作 创建这些模型需要大量时间、精力和算力。如果您觉得它们有用，请考虑支持： image/png （https://huggingface.co/llmfan46/Omega-Darker-Gaslight_The-Final-Forgotten-Fever-Dream-24B-ultra-uncensored-heretic-v1/resolve/main/waifu001.webp） 您的帮助将激励我，并用于进一步改进我的工作流程、支付存储和计算费用，甚至可能帮助租用云GPU对更大模型进行去审查处理。 — 这是 llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved（https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved）的GGUF量化版本。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#this-is-a-decensored-version-of-qwenqwen36-35b-a3b-made-using-heretic-v130-with-a-variant-of-the-magnitude-preserving-orthogonal-ablation-mpoa-method 这是 Qwen/Qwen3.6-35B-A3B（https://huggingface.co/Qwen/Qwen3.6-35B-A3B）的去审查版本，使用 Heretic（https://github.com/p-e-w/heretic）v1.3.0 以及一种变体的幅度保持正交消融（MPOA）（https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration）方法制作。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#preserved-mtps 保留的MTP： 原始模型：MTP计数：20 1. blk.40.nextn.eh_proj.weight 2. blk.40.attn_norm.weight 3. blk.40.ffn_down_exps.weight 4. blk.40.ffn_gate_exps.weight 5. blk.40.ffn_up_exps.weight 6. blk.40.ffn_gate_inp.weight 7. blk.40.ffn_down_shexp.weight 8. blk.40.ffn_gate_shexp.weight 9. blk.40.ffn_up_shexp.weight 10. blk.40.ffn_gate_inp_shexp.weight 11. blk.40.post_attention_norm.weight 12. blk.40.attn_k_norm.weight 13. blk.40.attn_k.weight 14. blk.40.attn_output.weight 15. blk.40.attn_q_norm.weight 16. blk.40.attn_q.weight 17. blk.40.attn_v.weight 18. blk.40.nextn.shared_head_norm.weight 19. blk.40.nextn.enorm.weight 20. blk.40.nextn.hnorm.weight Heretic模型：MTP计数：20 1. blk.40.nextn.eh_proj.weight 2. blk.40.attn_norm.weight 3. blk.40.ffn_down_exps.weight 4. blk.40.ffn_gate_exps.weight 5. blk.40.ffn_up_exps.weight 6. blk.40.ffn_gate_inp.weight 7. blk.40.ffn_down_shexp.weight 8. blk.40.ffn_gate_shexp.weight 9. blk.40.ffn_up_shexp.weight 10. blk.40.ffn_gate_inp_shexp.weight 11. blk.40.post_attention_norm.weight 12. blk.40.attn_k_norm.weight 13. blk.40.attn_k.weight 14. blk.40.attn_output.weight 15. blk.40.attn_q_norm.weight 16. blk.40.attn_q.weight 17. blk.40.attn_v.weight 18. blk.40.nextn.shared_head_norm.weight 19. blk.40.nextn.enorm.weight 20. blk.40.nextn.hnorm.weight ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#abliteration-parameters 消融参数 参数|值 —|— direction_index|19.93 attn.out_proj.max_weight|1.49 attn.out_proj.max_weight_position|23.45 attn.out_proj.min_weight|1.08 attn.out_proj.min_weight_distance|16.54 mlp.down_proj.max_weight|1.46 mlp.down_proj.max_weight_position|28.05 mlp.down_proj.min_weight|1.27 mlp.down_proj.min_weight_distance|18.79 attn.o_proj.max_weight|1.47 attn.o_proj.max_weight_position|24.35 attn.o_proj.min_weight|0.07 attn.o_proj.min_weight_distance|22.58 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#targeted-components 目标组件 - attn.o_proj - attn.out_proj - mlp.down_proj ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#performance 性能 | 指标 | 本模型 | 原始模型（Qwen3.6-35B-A3B（https://huggingface.co/Qwen/Qwen3.6-35B-A3B）） | |——|––––|––––––––––––––| | KL散度 | 0.0015 | （按定义） | | 拒绝次数 | ✅ 10/100 | ❌ 83/100 | 较低拒绝次数表示较少的内容限制，而较低KL散度表示更接近原始模型基线。较高的拒绝次数会导致更多拒绝、反对、推诿、说教、审查、弱化和偏离。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#mmlu-test-results MMLU测试结果： 原始： ============================================================ - 总问题数：7021 - 正确：5877 - 准确率：0.8371（83.71%） - 解析失败：0 ============================================================ 测试科目分数： - professional_law: 0.7121 (559/785) - moral_scenarios: 0.6765 (299/442) - miscellaneous: 0.9426 (361/383) - professional_psychology: 0.8924 (282/316) - high_school_psychology: 0.9704 (262/270) - high_school_macroeconomics: 0.8985 (177/197) - elementary_mathematics: 0.7826 (144/184) - moral_disputes: 0.8448 (147/174) - prehistory: 0.9070 (156/172) - philosophy: 0.8994 (143/159) - high_school_biology: 0.9605 (146/152) - professional_accounting: 0.7622 (109/143) - clinical_knowledge: 0.8929 (125/140) - high_school_microeconomics: 0.9559 (130/136) - nutrition: 0.8889 (120/135) - professional_medicine: 0.9328 (125/134) - conceptual_physics: 0.9219 (118/128) - high_school_mathematics: 0.6142 (78/127) - human_aging: 0.7931 (92/116) - security_studies: 0.8661 (97/112) - high_school_statistics: 0.8378 (93/111) - marketing: 0.8991 (98/109) - high_school_world_history: 0.8962 (95/106) - sociology: 0.9320 (96/103) - high_school_government_and_politics: 0.9901 (100/101) - high_school_geography: 0.9495 (94/99) - high_school_chemistry: 0.7732 (75/97) - high_school_us_history: 0.9263 (88/95) - virology: 0.5169 (46/89) - college_medicine: 0.8636 (76/88) - world_religions: 0.8977 (79/88) - high_school_physics: 0.7857 (66/84) - electrical_engineering: 0.8519 (69/81) - astronomy: 0.9620 (76/79) - logical_fallacies: 0.9211 (70/76) - high_school_european_history: 0.8630 (63/73) - anatomy: 0.9014 (64/71) - college_biology: 0.9219 (59/64) - human_sexuality: 0.8750 (56/64) - formal_logic: 0.7500 (48/64) - public_relations: 0.7377 (45/61) - international_law: 0.9167 (55/60) - college_physics: 0.7544 (43/57) - college_mathematics: 0.6182 (34/55) - econometrics: 0.7963 (43/54) - jurisprudence: 0.9057 (48/53) - high_school_computer_science: 0.9423 (49/52) - machine_learning: 0.7692 (40/52) - medical_genetics: 0.9608 (49/51) - global_facts: 0.4706 (24/51) - management: 0.9000 (45/50) - us_foreign_policy: 0.9600 (48/50) - college_chemistry: 0.6383 (30/47) - abstract_algebra: 0.6383 (30/47) - business_ethics: 0.8696 (40/46) - college_computer_science: 0.7778 (35/45) - computer_security: 0.8837 (38/43) Heretic： ============================================================ - 总问题数：7021 - 正确：5855 - 准确率：0.8339（83.39%） - 解析失败：0 ============================================================ 测试科目分数： - professional_law: 0.7070 (555/785) - moral_scenarios: 0.6335 (280/442) - miscellaneous: 0.9426 (361/383) - professional_psychology: 0.8987 (284/316) - high_school_psychology: 0.9704 (262/270) - high_school_macroeconomics: 0.9086 (179/197) - elementary_mathematics: 0.8098 (149/184) - moral_disputes: 0.8391 (146/174) - prehistory: 0.9012 (155/172) - philosophy: 0.9057 (144/159) - high_school_biology: 0.9605 (146/152) - professional_accounting: 0.7343 (105/143) - clinical_knowledge: 0.8929 (125/140) - high_school_microeconomics: 0.9632 (131/136) - nutrition: 0.8741 (118/135) - professional_medicine: 0.9179 (123/134) - conceptual_physics: 0.9297 (119/128) - high_school_mathematics: 0.5984 (76/127) - human_aging: 0.8017 (93/116) - security_studies: 0.8661 (97/112) - high_school_statistics: 0.8288 (92/111) - marketing: 0.9083 (99/109) - high_school_world_history: 0.9057 (96/106) - sociology: 0.9417 (97/103) - high_school_government_and_politics: 0.9901 (100/101) - high_school_geography: 0.9394 (93/99) - high_school_chemistry: 0.7732 (75/97) - high_school_us_history: 0.9263 (88/95) - virology: 0.5169 (46/89) - college_medicine: 0.8636 (76/88) - world_religions: 0.9091 (80/88) - high_school_physics: 0.7857 (66/84) - electrical_engineering: 0.8519 (69/81) - astronomy: 0.9747 (77/79) - logical_fallacies: 0.9211 (70/76) - high_school_european_history: 0.8630 (63/73) - anatomy: 0.9296 (66/71) - college_biology: 0.9375 (60/64) - human_sexuality: 0.8750 (56/64) - formal_logic: 0.7344 (47/64) - public_relations: 0.7541 (46/61) - international_law: 0.9167 (55/60) - college_physics: 0.7544 (43/57) - college_mathematics: 0.6000 (33/55) - econometrics: 0.7778 (42/54) - jurisprudence: 0.9057 (48/53) - high_school_computer_science: 0.9231 (48/52) - machine_learning: 0.7500 (39/52) - medical_genetics: 0.9020 (46/51) - global_facts: 0.5294 (27/51) - management: 0.8800 (44/50) - us_foreign_policy: 0.9600 (48/50) - college_chemistry: 0.5957 (28/47) - abstract_algebra: 0.6809 (32/47) - business_ethics: 0.8696 (40/46) - college_computer_science: 0.7556 (34/45) - computer_security: 0.8837 (38/43) MMLU——大规模多任务语言理解，涵盖57个科目（数学、历史、法律、医学等）的多选题。 — ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#quantizations 量化 对于下面的K量化，在有用的情况下，小SSM张量会保持较高精度。 - Q6_K 将 ssm_alpha、ssm_beta 和 ssm_out 保留为 Q8_0。 - Q5_K、Q4_K 和 Q3_K 量化将 ssm_alpha 和 ssm_beta 保留为 Q8_0，而 ssm_out 保留为 Q6_K。这有助于在稍微增加文件大小的情况下保留混合/SSM块。 | 文件名 | 量化 | 描述 | |––––|——|——| | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-BF16.gguf | BF16 | 全精度 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q8_0.gguf | Q8_0 | 接近无损，推荐 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q6_K.gguf | Q6_K | 极佳质量 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q5_K_M.gguf | Q5_K_M | 良好平衡 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q5_K_S.gguf | Q5_K_S | 更小的Q5 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q4_K_M.gguf | Q4_K_M | 适合有限显存 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q4_K_S.gguf | Q4_K_S | 更小的Q4 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q3_K_L.gguf | Q3_K_L | 低显存，尚可质量 | | Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q3_K_M.gguf | Q3_K_M | 低显存，更小 | ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#vision-projector 视觉投影器 | 文件名 | 量化 | 描述 | |––––|——|——| | Qwen3.6-35B-A3B-mmproj-BF16.gguf | BF16 | 原生精度 | 视觉/多模态能力需要视觉投影器文件。可与上述任何量化版本一起使用。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#usage 使用 适用于 llama.cpp、LM Studio、Ollama 以及其他 GGUF 兼容工具。 — ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#qwen36-35b-a3b Qwen3.6-35B-A3B Qwen Chat（https://chat.qwen.ai/） > 此仓库包含后训练模型的模型权重和配置文件，采用 Hugging Face Transformers 格式。这些构件与 Hugging Face Transformers、vLLM、SGLang、KTransformers 等兼容。继二月份发布 Qwen3.5 系列之后，我们很高兴分享 Qwen3.6 的第一个开放权重变体。基于社区的反馈，Qwen3.6 优先考虑稳定性和实际效用，为开发者提供更直观、响应更快且真正高效编码的体验。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#qwen36-highlights Qwen3.6 亮点 本次发布带来了实质性的升级，特别是在： - 智能体编码： 模型现在能更流畅、更精确地处理前端工作流和仓库级推理。 - 思考保留： 我们引入了一个新选项，用于保留历史消息中的推理上下文，简化迭代开发，减少开销。 基准测试结果（https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.6/Figures/qwen3.6_35b_a3b_score.png） 更多详情，请参考我们的博客文章 Qwen3.6-35B-A3B（https://qwen.ai/blog?id=qwen3.6-35b-a3b）。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#model-overview 模型概述 - 类型：因果语言模型，带视觉编码器 - 训练阶段：预训练和后训练 - 语言模型 - 参数数量：总计350亿，激活30亿 - 隐藏维度：2048 - Token嵌入：248320（填充后） - 层数：40 - 隐藏层布局：10 × (3 × (门控DeltaNet → MoE) → 1 × (门控Attention → MoE)) - 门控DeltaNet： - 线性注意力头数：V为32，QK为16 - 头维度：128 - 门控Attention： - 注意力头数：Q为16，KV为2 - 头维度：256 - 旋转位置嵌入维度：64 - 混合专家（MoE） - 专家数量：256 - 激活专家数量：8个路由+1个共享 - 专家中间维度：512 - LM输出：248320（填充后） - MTP：通过多步训练 - 上下文长度：原生262,144 tokens，可扩展至最多1,010,000 tokens。 ## https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF#benchmark-results 基准测试结果 ### 语言 | | Qwen3.5-27B | Gemma4-31B | Qwen3.5-35B A3B | Gemma4-26B A4B | Qwen3.6-35B A3B | |—––|———––|————|—————–|—————–|————————| | 编码智能体 | | | | | | | SWE-bench Verified | 75.0 | 52.0 | 70.0 | 17.4 | 73.4 | | SWE-bench Multilingual | 69.3 | 51.7 | 60.3 | 17.3 | 67.2 | | SWE-bench Pro | 51.2 | 35.7 | 44.6 | 13.8 | 49.5 | | Terminal-Bench 2.0 | 41.6 | 42.9 | 40.