Releases·Model Release
LuffyTheFox releases Qwen3.8-27B-Uncensored-Genesis-V1 GGUF using Marchenko-Pastur noise distillation
A community member released Qwen3.8-27B-Uncensored-Genesis-V1, a 27B fine-tune distributed in GGUF format (including an MTP variant) on Hugging Face. The fine-tune adapts ideas from an arXiv paper (1311.0851) to reduce excessive reasoning verbosity and token burn in LLMs.
93 engagement1 sourceWed, Sep 9, 2026, 02:09 PM
AI Overview
- Reddit user LuffyTheFox released Qwen3.8-27B-Uncensored-Genesis-V1-MTP-GGUF on Hugging Face on 2026-09-09.
- The model applies Marchenko-Pastur distribution-based distillation, per arXiv paper 1311.0851, to remove training noise the creator blames for verbose reasoning.
- LuffyTheFox asks for community testing, citing only an RTX 3060 with 12 GB VRAM; recommended settings include temperature=1.0 and top_p=0.95.
Related
Social·Community
Defiance: uncensored Qwen3.8-27B GGUF pairs HauhauCS weights with DavidAU's 40 AI-Construct persona modes
A community member released "Defiance", a GGUF build combining HauhauCS's uncensored Qwen3.8-27B weights byte-for-byte with DavidAU's Turbo-Brilliance chat template, enabling 40 selectable persona modes (Joker, Dr. House, 20-agent Einstein panel, Socratic tutor) without refusals or added caveats. Testing on an RTX 4090 showed 86-90 tok/s decode with 38-42% MTP draft-token acceptance. The build matters to local-model users who want uncensored, persona-driven inference that runs in llama.cpp, LM Studio, and Ollama out of the box.
A community member released "Defiance", a GGUF build combining HauhauCS's uncensored Qwen3.8-27B weights byte-for-byte with DavidAU's Turbo-Brilliance chat template, enabling 40 selectable persona modes (Joker, Dr. House, 20-agent Einstein panel, Socratic tutor) without refusals or added caveats. Testing on an RTX 4090 showed 86-90 tok/s decode with 38-42% MTP draft-token acceptance. The build matters to local-model users who want uncensored, persona-driven inference that runs in llama.cpp, LM Studio, and Ollama out of the box.
4 engagement1 sourceredditSat, Oct 17, 2026, 06:23 AM
Social·Community
User grafts MTP onto Qwen3.8-27B for low quants, saving RAM on Vulkan
A user created a Hugging Face model that grafts Multi-Token Prediction (MTP) onto Qwen3.8-27B, addressing the lack of MTP in Unsloth's lowest quantizations. The graft reduces RAM usage compared to using an external MTP file, but the model's quality is degraded when thinking is disabled.
A user created a Hugging Face model that grafts Multi-Token Prediction (MTP) onto Qwen3.8-27B, addressing the lack of MTP in Unsloth's lowest quantizations. The graft reduces RAM usage compared to using an external MTP file, but the model's quality is degraded when thinking is disabled.
6 engagement1 sourceredditSun, Aug 23, 2026, 08:27 AM
Social·Community
Community shares depth-pruned Qwen3.8-27B to ~22.7B params
A developer released Qwen3.8-23B-Mini-Me, a depth-pruned version of Qwen3.8-27B reduced to about 22.7B parameters by removing layers without fine-tuning. The model targets coding, agentic use, and multi-turn chats, with a smaller footprint but no benchmark claims.
A developer released Qwen3.8-23B-Mini-Me, a depth-pruned version of Qwen3.8-27B reduced to about 22.7B parameters by removing layers without fine-tuning. The model targets coding, agentic use, and multi-turn chats, with a smaller footprint but no benchmark claims.
662 engagement2 sourcesreddit, xWed, Aug 19, 2026, 11:19 PM
Social·Community
Users speculate Qwen 3.8 27B is a renamed Qwen 3.6 27B with more reasoning
A Reddit user running Qwen 3.8 27B on a single RTX 3090 with llama-server config suggests the model may actually be Qwen 3.6 27B with enhanced thinking capabilities. The post shares detailed configuration settings for local inference, sparking community discussion about the model's identity and reasoning behavior.
A Reddit user running Qwen 3.8 27B on a single RTX 3090 with llama-server config suggests the model may actually be Qwen 3.6 27B with enhanced thinking capabilities. The post shares detailed configuration settings for local inference, sparking community discussion about the model's identity and reasoning behavior.
12 engagement1 sourceredditFri, Aug 21, 2026, 02:28 PM
Releases·Model Release
Qwen 3.8 adds prompt-steered reasoning effort, but official template has bugs
Qwen released its first 3.8 model, introducing prompt-steered reasoning effort via a `reasoning_effort` parameter (xhigh, medium, low). However, the official Jinja chat template has serious issues: it crashes when thinking is disabled, poisons multi-turn chat history, and breaks tool calling with JSON string arguments. A community fix addresses these problems for Qwen 3.5, 3.6, and 3.8.
Qwen released its first 3.8 model, introducing prompt-steered reasoning effort via a `reasoning_effort` parameter (xhigh, medium, low). However, the official Jinja chat template has serious issues: it crashes when thinking is disabled, poisons multi-turn chat history, and breaks tool calling with JSON string arguments. A community fix addresses these problems for Qwen 3.5, 3.6, and 3.8.
Updated Sun, Aug 16, 2026, 01:42 PM — A second community template fixes issues in both the original and the first improved template.
012476 engagement4 sourcesredditThu, Aug 13, 2026, 08:22 PM