UkisAI has released the Swift family of efficient reasoning models, a set of Qwen-based LLMs designed to solve one specific problem that has plagued large reasoning models: they think too much. The company's previous model, Swift Qwen 3.8 27B, gathered over 350,000 downloads in thirteen days, and the new release extends the family with three additional variants that reduce token usage while maintaining or improving accuracy.

The Overthinking Problem

Large reasoning models are designed to think before they answer, which often produces better results than direct inference. The unintended side effect is that they can enter what UkisAI calls "overthinking error" loops, where the model pursues a task far beyond the point of diminishing returns, burning through tokens to chase marginal improvements or simply failing to converge on a solution. The result is a model that is slow, expensive, and sometimes less accurate than a smaller, more focused counterpart.

UkisAI's training approach targets this behavior directly. The models are trained by penalizing tokens associated with pathological overthinking patterns and restoring accuracy through reinforcement learning and Online Preference Distillation (OPD). The idea is to teach the model to recognize when it has enough information to answer and stop there.

The Three Models

The release includes three variants, each targeting a different tradeoff between reasoning depth and efficiency:

  • Swift1.5 27B: An improved version of the prior model that reduced thinking tokens by 58.5 percent while scoring 0.35 percent higher on benchmarks. It specifically outperformed the base model on Terminal Bench 2.1 by avoiding the overthinking trap, where models fail tasks not because they lack capability but because they loop endlessly.
  • Swift Flash Next: The most aggressive reduction in thinking tokens, cutting them by 63.4 percent with a 1.8x speed improvement while scoring within 0.2 percent of the base model on the xhigh benchmark tier.
  • Swift Bonsai 2: A more moderate reduction of 39.8 percent in thinking tokens with a 0.19 percent accuracy improvement, though the company marks this variant as experimental.

Benchmarking and the Terminal Bench Anomaly

All benchmarks were run five times across five seeds and multiple domains, including general reasoning (GPQA, AIME26), coding (LiveCodeBench), vision (ERQA), and agentic tasks (Terminal Bench 2.1). The company notes that Terminal Bench 2.1 scores appear misleadingly low at first glance. This is not a bug but a consequence of the Swift models not falling into overthinking loops. A base model that fails a task will often keep trying until it exhausts its token budget, accumulating a high average token count. The Swift model stops earlier, producing a lower average token count but also a lower apparent score on that particular benchmark. When compared on an apples-to-apples basis where failed tasks are treated equally, token reduction still falls in the 38.7 percent range.

Availability and Formats

UkisAI is providing a Free Research API and HuggingFace Spaces for users who want to test the models without downloading them locally. The full range of quantization formats is available, including GGUF, NVFP4, MLX, and W4A16, targeting different hardware configurations and deployment scenarios. The company operates what it describes as a "candy-shop basis" for fulfilling model requests and quantization needs, asking the community to share specific requirements in the comments.

The release addresses a genuine pain point for anyone running reasoning models in production. The cost of overthinking is not just theoretical; it shows up directly in inference bills and user-perceived latency. Whether the accuracy preservation at these reductions holds up under real-world workloads will depend on independent evaluations that the company has explicitly invited.