UkisAI
6 storiesUkisAI is a small AI lab that releases the Swift family of open-source reasoning models built on Qwen. Recent coverage has focused on those releases and their uptake: Swift-Qwen3.8-27B penalizes "overthinking" tokens to cut thinking tokens by 58% with a 1.95x speedup, Swift1.5-Qwen3.8-Flash-Next scored 41.1% first-try pass on the Aider benchmark while cutting tokens per case from 17,646 to 6,991, and the Swift series has passed 2.2 million downloads.
Related topics
Small lab UkisAI said its open-source Swift series of reasoning-efficient LLMs has passed 2.2 million downloads. By penalizing pathological overthinking, the models cut token usage 58.3% and boost speed 1.95x without losing accuracy. It also launched a Discord, with the first 100 members getting early access to unreleased Swift GLM 5.3 Flash and Swift 9B, plus UkisAI Code and Swift.cpp.
A user's comparison shows Swift-1.5-Qwen3.8-Flash-Next scored 41.1% first-try pass on the Aider benchmark versus 40.2% for the base model, while cutting tokens per case from 17,646 to 6,991 and time from 1,542s to 608s, sharply reducing overthinking.
UkisAI launched the Swift family of reasoning LLMs built on Qwen, cutting thinking tokens by penalizing pathological overthinking and restoring accuracy via RL. Swift Flash Next reduces thinking tokens 63.4% with 1.8x speedup; Swift1.5 27B cuts 58.5% with slightly higher scores; Swift Bonsai 2 cuts 39.8%. API, demos, and quants are provided.
UkisAI's open-source Swift Qwen 3.8 27B has surpassed 100,000 downloads on Hugging Face. The team says they reduced pathological overthinking to cut token use and speed up inference, and plan upcoming improved checkpoints and variants.
A LocalLLaMA post highlights UkisAI’s Swift‑Qwen3.8‑27B fine‑tune, which penalizes tokens that trigger overthinking and borrows BottleCap AI techniques to reduce reasoning tokens by roughly 30–50% while keeping output quality.
UkisAI post‑trained Qwen 3.8 27B to penalize tokens tied to “overthinking,” using On‑Policy Distillation to cut thinking tokens by 58%, speed up 1.95×, and keep accuracy loss under 1%. The model is open‑sourced on Hugging Face and a free NVIDIA‑backed OpenAI‑compatible API is available (5 RPM limit).
That is everything
The A.I. Daily
UkisAI's Swift open-source models surpass 2.2M downloads