TL;DR: UkisAI post-trained a 27B Qwen model to eliminate overthinking tokens, cutting reasoning length 58% and nearly doubling speed with under 1% accuracy loss.
Summary: UkisAI released Swift-Qwen3.8-27B, a post-trained variant that penalizes tokens associated with overthinking rather than directly capping reasoning length, then recovers any lost accuracy via on-policy distillation. The team reports a 58% reduction in thinking tokens and a 1.95x speedup while retaining the accuracy of the original high-effort setting, with under 1% degradation. The model is on Hugging Face with GGUF quants (Q1-Q8) and community NVFP4/W4A16 variants, plus a free OpenAI-compatible research API running on Nvidia-provided GPUs.
Why it matters: Cheaper, faster reasoning at near-identical accuracy directly lowers inference cost and latency for agentic and chain-of-thought workloads. Worth testing the GGUF quants locally or the free API against your own evals to see whether the token savings hold on your task distribution.
Source: reddit