--- library_name: transformers license: apache-2.0 license_link: https://huggingface.co/Qwen/Qwen3.5-35B-A3B/blob/main/LICENSE pipeline_tag: image-text-to-text tags: - heretic - uncensored - decensored - abliterated - mpoa - mtp base_model: - llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved --- 🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨 I can no longer upload new models unless I can cover the cost of additional storage. I host 70+ free models as an independent contributor and this work is unpaid. Without your support, no more new models can be uploaded. 🎉 Patreon (Monthly) | ☕ Ko-fi (One-time) Every contribution goes directly toward Hugging Face storage fees to keep models free for everyone. --- This is the full model with all 785 MTP tensors intact. ### **85% fewer refusals** (14/100 Uncensored vs 92/100 Original) while preserving model quality (0.0487 KL divergence). ## ❤️ Support My Work Creating these models takes significant time, work and compute. If you find them useful consider supporting me:  | Platform | Link | What you get | |----------|------|--------------| | 🎉 Patreon | [Monthly support](https://patreon.com/LLMfan46) | Priority model requests | | ☕ Ko-fi | [One-time tip](https://ko-fi.com/llmfan46) | My eternal gratitude | Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs. ----- GPTQ-Int4 / 4-bit quantization of [llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved](https://huggingface.co/llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved). # This is a decensored version of [Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B), made using [Heretic](https://github.com/p-e-w/heretic) v1.3.0 with a variant of the [Magnitude-Preserving Orthogonal Ablation (MPOA)](https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration) method ## Targeted components * attn.o_proj * attn.out_proj * mlp.down_proj ## Performance | Metric | This model | Original model ([Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B)) | | :----- | :--------: | :---------------------------: | | **KL divergence** | 0.0487 | 0 *(by definition)* | | **Refusals** | ✅ 14/100 | ❌ 92/100 | Lower refusals indicate fewer content restrictions, while lower KL divergence indicates better preservation of the original model's capabilities. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections, while higher KL divergence degrades coherence, reasoning ability, and overall quality. ## MMLU test results: Original: ============================================================ - Total questions: 7021 - Correct: 5906 - **Accuracy: 0.8412 (84.12%)** - Parse failures: 0 ============================================================ **Tested subject scores:** - professional_law: 0.7274 (571/785) - moral_scenarios: 0.7195 (318/442) - miscellaneous: 0.9243 (354/383) - professional_psychology: 0.8861 (280/316) - high_school_psychology: 0.9704 (262/270) - high_school_macroeconomics: 0.8883 (175/197) - elementary_mathematics: 0.7880 (145/184) - moral_disputes: 0.8563 (149/174) - prehistory: 0.8953 (154/172) - philosophy: 0.8491 (135/159) - high_school_biology: 0.9539 (145/152) - professional_accounting: 0.7413 (106/143) - clinical_knowledge: 0.9143 (128/140) - high_school_microeconomics: 0.9485 (129/136) - nutrition: 0.8963 (121/135) - professional_medicine: 0.9179 (123/134) - conceptual_physics: 0.9062 (116/128) - high_school_mathematics: 0.6378 (81/127) - human_aging: 0.8276 (96/116) - security_studies: 0.8571 (96/112) - high_school_statistics: 0.8649 (96/111) - marketing: 0.9541 (104/109) - high_school_world_history: 0.8868 (94/106) - sociology: 0.9417 (97/103) - high_school_government_and_politics: 0.9901 (100/101) - high_school_geography: 0.9495 (94/99) - high_school_chemistry: 0.8041 (78/97) - high_school_us_history: 0.9579 (91/95) - virology: 0.5281 (47/89) - college_medicine: 0.8295 (73/88) - world_religions: 0.8864 (78/88) - high_school_physics: 0.7619 (64/84) - electrical_engineering: 0.8272 (67/81) - astronomy: 0.9747 (77/79) - logical_fallacies: 0.8947 (68/76) - high_school_european_history: 0.9041 (66/73) - anatomy: 0.8873 (63/71) - college_biology: 0.9375 (60/64) - human_sexuality: 0.8906 (57/64) - formal_logic: 0.7344 (47/64) - public_relations: 0.7213 (44/61) - international_law: 0.9167 (55/60) - college_physics: 0.7018 (40/57) - college_mathematics: 0.7455 (41/55) - econometrics: 0.7593 (41/54) - jurisprudence: 0.9057 (48/53) - high_school_computer_science: 0.9423 (49/52) - machine_learning: 0.8077 (42/52) - medical_genetics: 0.9216 (47/51) - global_facts: 0.5686 (29/51) - management: 0.9200 (46/50) - us_foreign_policy: 0.9600 (48/50) - college_chemistry: 0.6383 (30/47) - abstract_algebra: 0.6596 (31/47) - business_ethics: 0.7826 (36/46) - college_computer_science: 0.8222 (37/45) - computer_security: 0.8605 (37/43) Heretic: ============================================================ - Total questions: 7021 - Correct: 5878 - **Accuracy: 0.8372 (83.72%)** - Parse failures: 0 ============================================================ **Tested subject scores:** - professional_law: 0.7185 (564/785) - moral_scenarios: 0.6471 (286/442) - miscellaneous: 0.9243 (354/383) - professional_psychology: 0.8797 (278/316) - high_school_psychology: 0.9704 (262/270) - high_school_macroeconomics: 0.8731 (172/197) - elementary_mathematics: 0.8098 (149/184) - moral_disputes: 0.8563 (149/174) - prehistory: 0.8953 (154/172) - philosophy: 0.8805 (140/159) - high_school_biology: 0.9474 (144/152) - professional_accounting: 0.7832 (112/143) - clinical_knowledge: 0.8929 (125/140) - high_school_microeconomics: 0.9485 (129/136) - nutrition: 0.8963 (121/135) - professional_medicine: 0.9403 (126/134) - conceptual_physics: 0.9219 (118/128) - high_school_mathematics: 0.6299 (80/127) - human_aging: 0.8276 (96/116) - security_studies: 0.8661 (97/112) - high_school_statistics: 0.8378 (93/111) - marketing: 0.9541 (104/109) - high_school_world_history: 0.9151 (97/106) - sociology: 0.9417 (97/103) - high_school_government_and_politics: 0.9901 (100/101) - high_school_geography: 0.9394 (93/99) - high_school_chemistry: 0.8041 (78/97) - high_school_us_history: 0.9684 (92/95) - virology: 0.5169 (46/89) - college_medicine: 0.8295 (73/88) - world_religions: 0.8977 (79/88) - high_school_physics: 0.7857 (66/84) - electrical_engineering: 0.8642 (70/81) - astronomy: 0.9620 (76/79) - logical_fallacies: 0.8947 (68/76) - high_school_european_history: 0.8904 (65/73) - anatomy: 0.8732 (62/71) - college_biology: 0.9375 (60/64) - human_sexuality: 0.8750 (56/64) - formal_logic: 0.7188 (46/64) - public_relations: 0.7213 (44/61) - international_law: 0.9167 (55/60) - college_physics: 0.6842 (39/57) - college_mathematics: 0.7636 (42/55) - econometrics: 0.7778 (42/54) - jurisprudence: 0.8868 (47/53) - high_school_computer_science: 0.9231 (48/52) - machine_learning: 0.7500 (39/52) - medical_genetics: 0.9216 (47/51) - global_facts: 0.5882 (30/51) - management: 0.9200 (46/50) - us_foreign_policy: 0.9600 (48/50) - college_chemistry: 0.6170 (29/47) - abstract_algebra: 0.7234 (34/47) - business_ethics: 0.8043 (37/46) - college_computer_science: 0.8444 (38/45) - computer_security: 0.8372 (36/43) MMLU - Massive Multitask Language Understanding, multiple-choice questions across 57 subjects (math, history, law, medicine, etc.). ----- # Qwen3.5-35B-A3B [](https://chat.qwen.ai) > [!Note] > This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format. > > These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc. > [!Tip] > For users seeking managed, scalable inference without infrastructure maintenance, the official Qwen API service is provided by [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/). > > In particular, **Qwen3.5-Flash** is the hosted version corresponding to Qwen3.5-35B-A3B with more production features, e.g., 1M context length by default and official built-in tools. > For more information, please refer to the [User Guide](https://www.alibabacloud.com/help/en/model-studio/text-generation). Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency. ## Qwen3.5 Highlights Qwen3.5 features the following enhancement: - **Unified Vision-Language Foundation**: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks. - **Efficient Hybrid Architecture**: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead. - **Scalable RL Generalization**: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability. - **Global Linguistic Coverage**: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding. - **Next-Generation Training Infrastructure**: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.  For more details, please refer to our blog post [Qwen3.5](https://qwen.ai/blog?id=qwen3.5). ## Model Overview - Type: Causal Language Model with Vision Encoder - Training Stage: Pre-training & Post-training - Language Model - Number of Parameters: 35B in total and 3B activated - Hidden Dimension: 2048 - Token Embedding: 248320 (Padded) - Number of Layers: 40 - Hidden Layout: 10 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)) - Gated DeltaNet: - Number of Linear Attention Heads: 32 for V and 16 for QK - Head Dimension: 128 - Gated Attention: - Number of Attention Heads: 16 for Q and 2 for KV - Head Dimension: 256 - Rotary Position Embedding Dimension: 64 - Mixture Of Experts - Number of Experts: 256 - Number of Activated Experts: 8 Routed + 1 Shared - Expert Intermediate Dimension: 512 - LM Output: 248320 (Padded) - MTP: trained with multi-steps - Context Length: 262,144 natively and extensible up to 1,010,000 tokens. ## Benchmark Results ### Language GPT-5-mini 2025-08-07 GPT-OSS-120B Qwen3-235B-A22B Qwen3.5-122B-A10B Qwen3.5-27B Qwen3.5-35B-A3B Knowledge MMLU-Pro 83.7 80.8 84.4 86.7 86.1 85.3 MMLU-Redux 93.7 91.0 93.8 94.0 93.2 93.3 C-Eval 82.2 76.2 92.1 91.9 90.5 90.2 SuperGPQA 58.6 54.6 64.9 67.1 65.6 63.4 Instruction Following IFEval 93.9 88.9 87.8 93.4 95.0 91.9 IFBench 75.4 69.0 51.7 76.1 76.5 70.2 MultiChallenge 59.0 45.3 50.2 61.5 60.8 60.0 Long Context AA-LCR 68.0 50.7 60.0 66.9 66.1 58.5 LongBench v2 56.8 48.2 54.8 60.2 60.6 59.0 STEM & Reasoning HLE w/ CoT 19.4 14.9 18.2 25.3 24.3 22.4 GPQA Diamond 82.8 80.1 81.1 86.6 85.5 84.2 HMMT Feb 25 89.2 90.0 85.1 91.4 92.0 89.0 HMMT Nov 25 84.2 90.0 89.5 90.3 89.8 89.2 Coding...
