Qwen3.6 27B uncensored K_P quants drop

This is useful for local model tinkerers, but the headline feature is also the risk: a "no refusals" derivative shifts responsibility for safety, evaluation, and deployment boundaries entirely onto the operator.

• –K_P quant support matters because 27B dense models live or die on quality-per-GB for consumer hardware users
• –The mmproj inclusion keeps the multimodal promise practical instead of leaving users to assemble vision support manually
• –Prompt sensitivity is the real tradeoff: weaker instruction clarity can mean more drift, especially in agent or coding workflows
• –The release rides the official Qwen3.6-27B launch, which gives it more technical relevance than a random fine-tune
• –Developers should treat the refusal benchmark claim as a community datapoint, not production validation