UkisAI Swift Series / 27B, Flash Next and Bonsai 2 + GSQ-RCO / -63.4% thinking, x1.95 speed with xhigh accuracy

Hey everyone, Jovan from UkisAI here! Today, we are introducing Swift, a family of efficient reasoning LLMs based on Qwen, trained by penalizing tokens related to pathological overthinking patterns and restoring accuracy via RL (GSPO) and OPD . After amazing feedback and 350k+ downloads in 13 days on our Swift Qwen 3.8 27B we are releasing the entire model family as well as the highly requested GSQ-RCO quants for 27B and Flash-Next. This release includes: Swift1.5 27B , an improved version of our last model, with even lower token usage, fixed bugs and better agentic performance, with -58.5% thinking tokens while scoring 0.35% higher and outperfoming base on Terminal Bench 2.1 by not falling into "overthinking error" loops. Swift Flash Next , with 63.4% fewer thinking tokens and a 1.8x speed up scoring -0.2% vs base on xhigh Swift Bonsai 2 , with 39.8% fewer thinking tokens while scoring 0.19% higher (although we'd still like to note it as experimental) Our benchmarks are ran x5 on Base and Swift, averaging across five seeds and various domains, including General (GPQA, AIME26), Coding (LiveCodeBench), Vision (ERQA), Agentic (Terminal Bench 2.1). One note is that the Terminal Bench 2.1 score of Swift1.5 27B is misleadingly low at first glance. It is not a bug, but a simple matter of the Swift models not falling into overthinking loops and failing the task, rather pursuing it until the end, leading to higher average token usage. The token reduction still falls in the -38.7% range when compared apples-to-apples. We also added a fun "game creation" benchmark you can find and play here, it is completely subjective but Swift generated better games in less time: Flash Next Game and 27B Game We are including a Research API and HuggingFace Spaces to give the models a spin before downloading or if you don't have enough compute to run them right now! You can find both on the model cards. We have also made GGUF, NVFP4, MLX and W4A16 quants for relevant model versions. More details on our training approach and community feedback can be seen here: All of the various quantization and model versions are available in their respective collections: Swift1.5 27B: huggingface.co/collections/ukisai/swift-15-27b Swift Flash Next: huggingface.co/collections/ukisai/swift-flash-next Swift Bonsai 2: huggingface.co/collections/ukisai/swift-bonsai-2 We are also working on a 9B variant to be released in the upcoming days. We would greatly appreciate your feedback via independent evaluations on real world tasks. As per last release, we operate on a candy-shop basis, trying to fulfill as many Swift model requests and quants as possible, so please do share your needs in the comments!