The fine-tune Qwen 3.8 27B with reasoning optimization has been released

Developers presented Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored — a specialized version of the model focused on high performance and removal of censorship restrictions.

The key feature of the release was architectural optimization, allowing to reduce the number of tokens spent on the "thinking" (reasoning) process by up to 95%. At the same time, developers claim to preserve intellectual abilities at the level of OpenAI and Gemini models, which is confirmed by the Arc-C benchmark scores at the level of 699/692 points.

The model supports 5 switchable modes of operation, including specialized configurations for reasoning and instruction execution. The build is presented in the GGUF format, which provides compatibility with a wide range of local tools for running LLM.

This project is an example of a deep refinement of the basic Qwen architecture for specific tasks, where the speed of generating answers while maintaining logical coherence is critically important. The use of Twin-Turbo optimization allows to significantly reduce computational costs for generating long chains of reasoning, making the model suitable for use in resource-intensive scenarios with limited response time.