Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-FP8-Dynamic

Qwen3.6-27B Uncensored HauhauCS Aggressive — FP8-Dynamic

vLLM-native FP8 quantization (compressed-tensors, data-free dynamic scheme) of the HauhauCS uncensored finetune.

Provenance & fidelity: HauhauCS released this finetune only as GGUF — no safetensors source exists publicly. These weights were dequantized from the highest-fidelity public artifact (Q8_K_P GGUF, which internally is Q8_0/F16/F32 tensors) back into the official Qwen/Qwen3.6-27B HF layout, with every tensor validated against the stock checkpoint's exact shape and dtype (1199/1199). Quality ceiling is therefore Q8 (near-lossless vs the author's private BF16), not true BF16.

• Vision tower included (rebuilt from the author's f16 mmproj; qkv re-fused, Conv3D patch-embed re-stacked) — full image/video input support.
• MTP head included (mtp.*, taken from stock Qwen/Qwen3.6-27B since the finetune never shipped one) — enables speculative decoding in engines that support Qwen3.6 MTP.
• Tokenizer/configs from stock Qwen/Qwen3.6-27B (the finetune does not modify them).

Usage (vLLM)

vllm serve zeeksa/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-FP8-Dynamic

compressed-tensors FP8_DYNAMIC (float-quantized weights, dynamic per-token activations). Linear layers FP8; lm_head and the vision tower kept BF16. ~31 GB.

Related

• GGUF + DFlash speculative decoding bundle: zeeksa/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-DFlash-GGUF
• Original finetune (GGUF only): HauhauCS/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive (Apache-2.0)

Note: Qwen3.6's gated attention is new — use a recent transformers (>=5.10) / recent vLLM; older engine versions may not support the architecture.