Swift-Qwen3.8-27B Cuts Reasoning Tokens Without Sacrificing Much Accuracy

UkisAI introduced Swift-Qwen3.8-27B, a derivative of Qwen3.8-27B with 27B parameters and 262,144-token context. The model reduces median reasoning tokens by 58.3% on GPQA-Diamond with a 0.1 pp accuracy drop, speeding up to 1.95x.

• 27B parameters, 262,144-token context, runs via Transformers, vLLM, or SGLang
• GPQA-Diamond: 88.28% vs 88.38% base, tokens reduced from 15,014 to 8,855
• LiveCodeBench v6: 81.55% vs 76.76%, average tokens fell from 11,374 to 8,615
• Swift Open License v1.0 free for revenue up to $1M per year