Qwen3.5-27B-Claude-4.6-Opus-Uncensored-V2-Kullback-Leibler-GGUF
An uncensored hybrid model now holds 262K context and reasons like Claude Opus 4.6, all without additional training.
A new open-source AI model, 'Qwen3.5-27B-Claude-4.6-Opus-Uncensored-V2', demonstrates the power of community-driven model merging. Built by developer LuffyTheFox, it starts with Alibaba's capable Qwen2.5-27B-Instruct model. This base was then fine-tuned on a 10,000-sample dataset curated by Roman1111111, designed to replicate the reasoning patterns and outputs of Anthropic's top-tier Claude Opus 4.6 model. The result is a hybrid that achieves a remarkable 96.91% pass rate on the HumanEval coding benchmark, rivaling the performance of much larger proprietary models.
Key technical work involved 'uncensoring' the model using the HauhauCS technique and fixing a critical Kullback-Leibler (KL) divergence issue, reducing the parametric KL from 1.14 to 0.28—a 75.6% improvement that stabilizes outputs. The developer also restored broken attention and feed-forward network layers during the conversion to the efficient GGUF format, enabling a massive 262,000-token context window. The model is available in quantized versions (like Q4_K_M and Q8_0) for local deployment, though its dense 27-billion-parameter size makes it slow on consumer hardware like an RTX 3060, highlighting the ongoing trade-off between capability and accessibility in the open-source AI space.
- Achieves 96.91% on HumanEval benchmark by merging Qwen2.5-27B with a Claude Opus 4.6-style dataset.
- Holds 262K context window with a 75.6% reduction in parametric KL divergence for more stable outputs.
- Available as quantized GGUF files for local use, but runs slowly (4 tokens/sec) on a 12GB RTX 3060 GPU.
Why It Matters
Shows how open-source communities can remix and enhance state-of-the-art models, pushing the boundaries of local, uncensored AI.