Huihui Qwen3.6 27B abliterated AWQ AWQ W4A16 quantized version of huihui ai/Huihui Qwen3.6 27B abliterated. This repository is marked as a quantized derivative of the Huihui model via: Quantization The model uses native AutoAWQ style AWQ INT4 weights with FP16 activations: Additional modules intentionally left unquantized are recorded in config.json under quantization config.modules to not convert . Tested Runtime Validated locally with a modified 1Cat vLLM build on 4 x Tesla V100 SXM2 32GB: The tested local server used SM70 AWQ kernels, FLASH ATTN V100 , and FP8 KV cache. For contexts above the model config limit, vLLM requires VLLM ALLOW LONG MAX MODEL LEN=1 ; use that override only after validating quality/stability for your workload. Notes This model inherits the safety/usage characteristics of the upstream abliterated model. The upstream authors describe it as an uncensored/abliterated variant of Qwen3.6 27B and warn that safety filtering is reduced. Review outputs before using in production or public facing systems. Base Model Quantized from: huihui ai/Huihui Qwen3.6 27B abliterated Original base model referenced by upstream: Qwen/Qwen3.6 27B
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy