UkisAI Releases Swift Series Models Reducing Reasoning Tokens by Up to 63%

UkisAI has released the Swift model family based on Qwen—including Swift1.5 27B, Swift Flash Next, and Swift Bonsai 2—which cuts overthinking reasoning tokens by up to 63.4% while accelerating generation speeds by up to 1.95x. The models were trained by penalizing unnecessary thinking patterns using Group Sequence Policy Optimization (GSPO) and On-Policy Distillation (OPD) without degrading benchmark accuracy. Modern reasoning LLMs often produce excessive intermediate thinking tokens, increasing latency, compute costs, and the likelihood of falling into repetitive error loops. By significantly reducing token bloat while maintaining output quality, the Swift series makes local deployment of high-performance reasoning models faster and far more cost-effective. The release features Swift Flash Next (-63.4% thinking tokens, 1.8x speedup) and Swift1.5 27B (-58.5% thinking tokens), along with multiple quantization formats including GSQ-RCO, GGUF, NVFP4, MLX, and W4A16. In agentic evaluations like Terminal Bench 2.1, Swift models avoid getting stuck in overthinking loops, allowing them to pursue tasks to completion rather than failing early.