UkisAI
A small AI laboratory focused on open-source large language model research and optimization, with a commitment to improving the reasoning efficiency of smaller LLMs.
Timeline (last 90 days)
The Swift series of models is trained by the UkisAI team using reinforcement learning, applying a penalty signal to the chain-of-thought length of reasoning models in order to shorten internal reasoning steps.
Swift Qwen3.8 27B downloads have surpassed 100,000.
Swift Qwen3.8 27B is UkisAI's first open-source model.
The approach adopted by UkisAI does not directly train the model to shorten its thinking, but instead trains it to think more efficiently, in order to address the overthinking problem in small LLMs.
UkisAI is advancing Swift1.5 Qwen3.8 27B, fixing bugs from training and incorporating more reinforcement learning training.
Swift Qwen3.8 Flash Next is expected to be released within the next week.
The team restored accuracy loss to within 1% in stages through reinforcement learning (GSPO), On-Policy Distillation, and a ThinkingCap 3.6 27B adapter, treating On-Policy Distillation as the key step.
UkisAI has open-sourced the Swift-Qwen3.8-27B model on Reddit.
The team is developing Swift 3.8 Flash Next, which has already achieved -30% thinking tokens while maintaining xhigh accuracy
The team has opened an OpenAI-compatible free research-use API with support from Nvidia GPUs, rate-limited to 5 RPM
5 more timeline events
All Facts (12)
The Swift series of models is trained by the UkisAI team using reinforcement learning, applying a penalty signal to the chain-of-thought length of reasoning models in order to shorten internal reasoning steps.
50%UnverifiedThe approach adopted by UkisAI does not directly train the model to shorten its thinking, but instead trains it to think more efficiently, in order to address the overthinking problem in small LLMs.
50%UnverifiedSwift Qwen3.8 27B is UkisAI's first open-source model.
50%UnverifiedUkisAI's methodology draws on a paper by Meta and uses an 8×H100 cluster to generate reasoning trajectories in out-of-distribution domains including coding, language, vision, and agents.
50%UnverifiedThe team built an inference-time penalizer that outperforms the original token set provided in the paper and is effective for both low-precision and BF16 models.
50%UnverifiedThe team used the identified tokens to construct a loss function and ran LoRA SFT, successfully reducing reasoning length but causing accuracy to drop.
50%UnverifiedThe team restored accuracy loss to within 1% in stages through reinforcement learning (GSPO), On-Policy Distillation, and a ThinkingCap 3.6 27B adapter, treating On-Policy Distillation as the key step.
50%UnverifiedThe UkisAI team believes that reasoning length itself is important and should not be forcibly shortened but rather optimized, and that their approach is a complement to, not a replacement for, existing reasoning effort mechanisms.
50%UnverifiedOverthinking errors are also prevalent in same-sized models at BF16 full precision, and are not merely a side effect of low-precision quantization (PTQ).
50%UnverifiedUkisAI is advancing Swift1.5 Qwen3.8 27B, fixing bugs from training and incorporating more reinforcement learning training.
50%UnverifiedThe team is developing Swift 3.8 Flash Next, which has already achieved -30% thinking tokens while maintaining xhigh accuracy
50%UnverifiedUkisAI has open-sourced the Swift-Qwen3.8-27B model on Reddit.
50%