Benchmark: ThinkingCap vs Swift vs Qwen 3.8-27B

A third-party Aider evaluation found that ThinkingCap and Swift fine-tunes cut Qwen3.8-27B's median reasoning tokens from 12,547 to about 7,400, a roughly 40% reduction matching their model cards, with first-try pass rates of 27.1%-30.8% and retry pass rates of 75.7%-77.6%, showing no meaningful performance loss.

• Swift 1.5 27b released: Swift Qwen gets faster2026-09-26
• UkisAI Releases Swift Family of Efficient Reasoning Models2026-09-24
• Yandex Releases 80B AliceAI Foundation Model2026-09-22
• Qwen 27B runs autonomously for 3 weeks on one 30902026-09-21
• Ling 3.0 Tiny brings edge intelligence to old laptops2026-09-26