Open-source Swift-Qwen3.8-27B cuts thinking tokens 58% for 1.95x speed at <1% accuracy loss
Secure_Recording_472 · reddit · 2026-09-14
UkisAI open-sourced Swift-Qwen3.8-27B, a post-trained Qwen 3.8 27B that penalizes tokens linked to overthinking loops (identified by clustering OOD traces across coding/language/vision/agentic domains) without force-shortening reasoning, then restores accuracy via on-policy distillation and RL. Results: -58% thinking tokens, 1.95x speedup, <1% accuracy loss. GGUF quants, community variants, and a free Nvidia-backed research API (5 RPM) are available on Hugging Face.
Related event: Open-Source Swift-Qwen3.8-27B Tops Hugging Face Trending(3 posts)→
More from Models
- Your AI agent would cost 260% more without prompt caching, Arena data shows — arena · 2026-10-30
- EachLabs launches Lockstep recast model claiming 75% cost cut and better quality — miilesus · 2026-10-30
- ChatGPT cited a 300+ page government legal PDF it retrieved on its own — lilyraynyc · 2026-10-30
- Best MLX model for French-to-English translation on an M2 Mac? — greenreddits · 2026-10-30
- AI-driven cyber ops near the sophistication peak set by GTG-1002's Claude Code campaign — Jsevillamol · 2026-10-30
- Cognition FrontierCode chart pits Mistral Large 4 vs Chinese open models, sparking debate — onetwoval · 2026-10-30