UkisAI announces Swift GLM-5.3-Flash and Swift Ornith 9B, two reasoning-efficient models that cut thinking token usage significantly (-61.4% and -33.5% respectively) while preserving accuracy, using GSQ-RCO penalization of pathological overthinking tokens, RL-based adaptation, and on-policy distillation, with multiple quantization formats released.
Hi everyone, Jovan from UkisAI here, Today we are expanding our Swift model series with two models requested by the community. We made Swift GLM-5.3-Flash and Swift Ornith 9B more efficient by figuring out which tokens were linked to pathological overthinking and penalizing them without "attacking" the reasoning length directly then regained the accuracy with further post-training, leading to "compressed" token usage while maintaining accuracy by making the model adapt to less reasoning via RL and on-policy distillation. You can find our dataset here (albeit we re-sample it per model) and more about our training approach here. Swift1.5 GLM-5.3-Flash, with the headline number being -61.4% thinking tokens (on MMLU-PRO) and -0.46% accuracy loss average on all the benchmarks we measured, all ran x5 on different seeds (Terminal-Bench 2.1, DeepSWE v1.1, GPQA-Diamond, MMLU-Pro, C-Eval, IFBench, AIME 2026, HMMT, LiveCodeBench v6) For this model, we are releasing a comprehensive number of quants, a GGUF(1bit-8bit), W4A16AWQ, NVFP4 and are also working on independently replicating GSQ-RCO papers, with our research so far showing our Q3 GSQ-RCO to be SOTA at it's size, meaning that we will be releasing it in the upcoming days after our evaluation suite is done. GGUF with 1bit-8bit quants, https://huggingface.co/ukisai/Swift-1.5-GLM-5.3-Flash-GGUF & a community favorite (at least with the folk with less VRAM) Swift1.5 Ornith-9B, with -33.5% thinking token usage with no accuracy loss, but more importantly ~2x less overthinking loops, meaning when you prompt the model it's much less likely for it to loop and not finish the prompt you gave it. Note that this is an experimental release and we are open to feedback as the model is already post-trained before our token optimizations. GGUF with 4bit-8bit quants, https://huggingface.co/ukisai/Swift-1.5-Ornith-9B-GGUF Both models work with their base MTP or DFLASH2 heads. We kindly request the community to let us know what other models we need to make more reasoning efficient and also if there are any other quants you want us to make. A little sneak peak is that we are currently evaluating Swift DeepSeek v4.1 Flash 1017 and it seems to be our best model yet! We are also cooking our own foundation, with the 1B and 4B models expected to release Nov. 2026., they are pre-trained with a diverse in-house coding dataset and post-trained to be token efficient like our existing Swift models. Hope you all like this release! And as always, all feedback is welcome and I'll try to respond to each request of the community :)
UkisAI releases the Swift Series of efficient reasoning LLMs based on Qwen, with models offering significant token reduction and accuracy improvements, along with various quantization options for deployment.
UkisAI has post-trained the Qwen 3.8 27B model to reduce unnecessary thinking tokens by 58% and achieve 1.95x speed up with less than 1% accuracy loss, open-sourcing the model and offering a free research API.
Swift-Qwen3.8-27B is UkisAI's reasoning-efficient derivative of Qwen3.8-27B, reducing thinking tokens by 58.3% while maintaining near-identical performance.
Release of Swift 1.5 Qwen3.8-27B mixed-precision GGUF quantizations, refined using GSQ-RCO per-tensor allocations from ISTA-DASLab. Swift 1.5 reduces thinking tokens by 58.5% with a 9.18× speed-up while maintaining or slightly improving benchmark scores.
Ornith-1.5-35B-A3B is a new AI foundation model that achieves superior performance on coding and agentic benchmarks by employing end-to-end self-improvement, activating only about 3 billion parameters per token.