Swift-1.5-Qwen3.8-Flash-Next
A local inference model trained by the UkisAI team using reinforcement learning, built on Qwen3.8 as its base. By applying penalty signals to Chain-of-Thought length, the model substantially shortens reasoning steps without sacrificing answer accuracy, resulting in reduced token consumption and lower inference latency.
Timeline (last 90 days)
While maintaining nearly equivalent quality, the Swift version compresses both reasoning token consumption and latency to approximately 40% of the base version, Qwen3.8-Flash-Next.
In the Aider benchmark, Swift-1.5-Flash-Next (xhigh) achieved a first-attempt pass rate of 41.1%, slightly above Qwen3.8-Flash-Next's 40.2%.
The tester concluded that only a difference of approximately 5% or more would be statistically meaningful on a dataset of this size, and since the actual gap is under 4%, it has not reached a significant level.
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's optimization occurs at the level of reasoning behavior with no change to the model's parameter count, which is fundamentally different from distillation or quantization compression methods that target model weights.
In a comparison of the 20 runs where the base version consumed the most tokens, Swift used only 29% of the tokens yet solved 16 problems, nearly matching the base version's 17.
Swift used at most 44K tokens on any single case, while the base version reached a maximum of 203K tokens.
In C++ language tests, Swift's token usage was only 29% of the base version's, but accounted for 3 of the 6 quality losses.
For heavy C++ developers, Swift's aggressive reasoning compression introduces quality fluctuations, and it is recommended to moderately relax the reasoning budget for greater stability.
Paired statistics (n=107) show 99 agreements, 2 cases where Swift was better, and 6 cases where it was worse; McNemar's exact test yielded p≈0.29, making the quality of the two statistically indistinguishable.
2 more timeline events
All Facts (12)
While maintaining nearly equivalent quality, the Swift version compresses both reasoning token consumption and latency to approximately 40% of the base version, Qwen3.8-Flash-Next.
50%UnverifiedIn the Aider benchmark, Swift-1.5-Flash-Next (xhigh) achieved a first-attempt pass rate of 41.1%, slightly above Qwen3.8-Flash-Next's 40.2%.
50%UnverifiedThe tester concluded that only a difference of approximately 5% or more would be statistically meaningful on a dataset of this size, and since the actual gap is under 4%, it has not reached a significant level.
50%UnverifiedThe 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%UnverifiedSwift's optimization occurs at the level of reasoning behavior with no change to the model's parameter count, which is fundamentally different from distillation or quantization compression methods that target model weights.
50%UnverifiedIn a comparison of the 20 runs where the base version consumed the most tokens, Swift used only 29% of the tokens yet solved 16 problems, nearly matching the base version's 17.
50%UnverifiedSwift used at most 44K tokens on any single case, while the base version reached a maximum of 203K tokens.
50%UnverifiedIn C++ language tests, Swift's token usage was only 29% of the base version's, but accounted for 3 of the 6 quality losses.
50%UnverifiedFor heavy C++ developers, Swift's aggressive reasoning compression introduces quality fluctuations, and it is recommended to moderately relax the reasoning budget for greater stability.
50%UnverifiedPaired statistics (n=107) show 99 agreements, 2 cases where Swift was better, and 6 cases where it was worse; McNemar's exact test yielded p≈0.29, making the quality of the two statistically indistinguishable.
50%UnverifiedThe test used the Swift version's Q5_K_L and the base version's Unsloth Q5_K_XL quantization configurations, both running in 128GB of memory with support for a 262K context, with engrams in Q8_0 stored on SSD.
50%UnverifiedSwift's valid diff rate reached 100%, meaning the generated code patch format was fully compliant.
50%