Qwen3.5-122B-A10B Uncensored (Aggressive) — GGUF Release + new K_P Quants
A 122B parameter AI model with zero refusals and new 'K_P' quants that boost quality by 1-2 levels.
Independent developer HauhauCS has released a significant uncensored version of Alibaba's Qwen3.5 model, dubbed the 'Qwen3.5-122B-A10B Aggressive.' This release strips away all refusal mechanisms from the original 122-billion-parameter model, resulting in a system that answers any prompt without capability loss, as evidenced by its 0/465 refusal score. The model is a Mixture-of-Experts (MoE) architecture with 256 experts, 8+1 active per token, a 262K context window, and multimodal (text+image+video) support. It's designed for local deployment in GGUF format, compatible with llama.cpp and LM Studio.
A major technical innovation in this release is the introduction of new 'K_P' ('Perfect') quantizations. Developed through model-specific analysis, these quants selectively preserve quality in critical areas, offering a 1-2 quantization level improvement for only a 5-15% increase in file size. For example, a Q4_K_P quant performs closer to a standard Q6_K. The release includes a full suite of quants from Q8_K_P down to IQ2_M, all generated with an importance matrix (imatrix) for optimal quality. The developer notes that while the aggressive uncensoring was 'brutal' work, the model shows no looping or degradation in testing, though users must disable its 'thinking' mode via a jinja template or API flag.
- Model is fully uncensored with 0/465 refusals, removing all refusal mechanisms from Alibaba's original Qwen3.5-122B-A10B.
- Introduces new 'K_P' quantizations that improve quality by 1-2 levels (e.g., Q4_K_P performs like Q6_K) for only a 5-15% file size increase.
- A 122B parameter Mixture-of-Experts model with 262K context, multimodal capabilities, and full local deployment via GGUF format.
Why It Matters
Enables unfiltered, high-capacity AI experimentation and application locally, pushing the boundaries of open-weight model customization and efficiency.