Qwen3 uncensored checkpoints land in 4B and 80B sizes
User 2etatg released four uncensored Qwen3 checkpoints on May 18, spanning 4B and 80B parameter counts with GRPO and SFT training variants, plus a GGUF quantized version for local inference.
Four uncensored Qwen3 checkpoints appeared on HuggingFace on May 18, 2026, covering 4B and 80B parameter scales. The releases include Qwen3-4B-Thinking-2507-GRPO-Uncensored-V2, a 4-billion-parameter model trained with group-relative policy optimization, and three 80B variants: a GRPO-trained checkpoint, an SFT-trained checkpoint, and a GGUF-quantized version of the GRPO weights. All four support English, Chinese, and Japanese text generation and ship as safetensors or GGUF files.
The GGUF quantized 80B variant has logged 41 downloads to date, compared to 16 for the 4B GRPO checkpoint, 14 for the 80B GRPO checkpoint, and 10 for the 80B SFT variant. GGUF quantization compresses weights for CPU and consumer-GPU inference, making the 80B scale accessible without datacenter hardware—a standard distribution path for uncensored variants in the local-inference community.