DeepSeek-Qwen3.5 9B Hits 90.53% MMLU-Pro Running Uncensored on Consumer Hardware
A 9B reasoning model distilled from DeepSeek-V4-Pro and decensored with Heretic now ships as imatrix GGUFs that run on 4GB GPUs.
Takeaways
- Mradermacher published imatrix GGUF quants of DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic, a decensored Qwen3.5-9B reasoning fine-tune.
- Base model distilled from DeepSeek-V4-Pro Max Effect traces, then stripped of refusal behavior using Heretic v2.0.0.dev0.
- Benchmarks: 94.50% GSM8K and 90.53% MMLU-Pro average across math, physics, and chemistry.
- Quant ladder spans IQ1_S at 2.8 GB up to Q6_K at 7.5 GB, with Q4_K_M flagged as the recommended default.
- Runs on 4GB VRAM GPUs with IQ1_M or IQ2_M, including vision via a 1.12 GB mmproj file.
- Apache 2.0 licensed and compatible with llama.cpp, Ollama, LM Studio, Jan, and Docker Model Runner.
Weighted importance-matrix GGUF quantizations are now available for the source checkpoint, a 9-billion-parameter Qwen3.5 derivative trained on DeepSeek-generated reasoning traces and modified to reduce refusals. The quantized release offers roughly two dozen sizes for llama.cpp-compatible runtimes. At publication, Hugging Face displayed more than 32,000 downloads for the repository.
Three stages, one checkpoint
The release combines work from several projects, so its name describes a lineage rather than a single model architecture:
- Distillation: Jackrong’s DeepSeek-V4-Pro-Qwen3.5-9B fine-tune uses responses generated by DeepSeek-V4-Pro in “Max Effect” mode. The 9B student learns from outputs produced with the teacher’s higher-compute reasoning configuration.
- Refusal suppression: Heretic v2.0.0.dev0 applies post-training weight edits intended to reduce refusal behavior without a conventional full fine-tune.
- Quantization: mradermacher converts the resulting checkpoint into GGUF files using importance-matrix calibration, which preserves more precision for weights that calibration data identifies as sensitive.
The underlying architecture is Qwen3.5-9B. The model cards list vision, tool-use and function-calling capabilities, although runtime support depends on the inference engine and chat template. Vision requires a separate multimodal projector from the