Qwen3.5 4B fits uncensored roleplay

This thread is a good reminder that local LLM "uncensored" requests are mostly hardware-fitting problems, not secret-model-hunting problems. For an 8GB card, the 4B Qwen variant is the first suggestion that looks genuinely sane.

• –The model card lists GGUF quants as low as 2.6GB for Q4_K_M and 3.3GB for Q6_K, so it leaves real headroom for context and runtime overhead on a 3060 Ti.
• –It is natively multimodal with 262K context, which is more capability than this use case strictly needs, but a nice bonus if the user wants image inputs later.
• –The "aggressive" uncensored variant is explicitly tuned to strip refusals, which suits roleplay but also means looser guardrails by design.
• –Bigger Qwen3.5-35B-class suggestions in the thread are more aspirational than practical unless the user accepts a lot of offload and latency.