Uncensored Qwen3.8-27B GGUF variants dominate Hugging Face trending

From August 15 to 17, multiple Qwen3.8-27B derivatives with safety filters removed flooded the Hugging Face trending chart: JonathanColetti's Uncensored-GGUF version, Blackfrost-AI's ABLITERATED-GGUF version, and the Heretic-Abliterated-Uncensored version reported by 0bserverx gained traction one after another, while Unsloth's GGUF quantized version racked up 1,000 likes in under 24 hours and reached 1 million total downloads. This wave is driven jointly by abliteration (de-censoring) and low-barrier quantized deployment, and warrants continued attention.

Confirmed

• Three main versions: Qwen3.8-27B-Uncensored-GGUF (JonathanColetti, 08-15, supports llama.cpp inference), Qwen3.8-27B-ABLITERATED-GGUF (Blackfrost-AI, 08-16, a 27-billion-parameter dense model claimed to support image-text-to-text multimodal capability), and Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF (reported by 0bserverx, launched 08-17).
• Common approach: abliteration strips the original model's safety filters and refusal mechanisms to enable unrestricted text generation; the GGUF quantized format is the mainstream release format.
• Unsloth's Qwen3.8-27B GGUF quantized version earned 1,000 likes in under 24 hours, ranking #3 on the HF trending chart with 1 million total downloads, and runs on a 17GB RAM/VRAM setup (as relayed by danielhanchen).
• User feedback: Frizzy-MacDrizzle reports the Ablit model has a low refusal rate in their analytics service and performs well.

Unconfirmed

• 0bserverx notes the Heretic version was mislabeled as 3.5 and believes it is actually based on Qwen 3.8; the mapping between each version and the official base model should be verified against the release pages.
• The multimodal capability is described only by Blackfrost-AI and is not corroborated by other posts.

Why It Matters

• De-censored models keep dominating the charts, reflecting strong community demand for locally deployed, unrestricted generative models.
• A 17GB RAM/VRAM footprint means consumer-grade hardware can run a 27B-class model, with the quantization ecosystem around Unsloth, GGUF, and llama.cpp acting as a key driver of this spread.

2026-08-15 ~ 2026-08-17 · 6 related posts

Full story(4 episodes)→

• Episode 1: Alibaba's Open-Source Qwen 3.8 27B Draws Rave Local Reviews, Nears Claude Opus(2026-08-15, 6 posts)
• Episode 2: Alibaba Open-Sources Qwen3.8-27B Multimodal Model That Runs on a Single GPU(2026-08-15, 2 posts)
• Episode 3: Uncensored Qwen3.8-27B GGUF variants dominate Hugging Face trending(2026-08-15, 6 posts)
• Episode 4: Qwen3.8-27B Gets INT4 and MXFP4 Quantized Releases(2026-08-16, 2 posts)

• Qwen3.8-27B GGUF Hits 1,000 Likes in 24 Hours — danielhanchen · 2026-08-15
• [source] Uncensored Qwen3.8-27B GGUF Model Trends on Hugging Face — JonathanColetti · 2026-08-15
• Uncensored Qwen 27B Model Released as GGUF Quantization — BLUECOW009 · 2026-08-16
• [source] Multimodal Qwen3.8-27B-ABLITERATED-GGUF Trends on Hugging Face — Blackfrost-AI · 2026-08-16
• Qwen 3.8 Ablit model available on Hugging Face — Frizzy-MacDrizzle · 2026-08-16
• [source] Uncensored Qwen 3.8 27B model released on Hugging Face — 0bserverx · 2026-08-17