Swift-Qwen3.8-27B cuts reasoning tokens with a lightweight adapter

Reasoning models often generate thousands of tokens while revisiting intermediate steps before producing an answer. UkisAI’s product page describes Swift-Qwen3.8-27B, a fine-tune designed to shorten those reasoning traces. UkisAI reports 58.3% fewer thinking tokens, less than 1% aggregate performance loss, and a 1.95x speedup on several measured tasks.

Swift is distributed through adapter weights for Qwen3.8-27B. The LoRA-style adapter adds a small set of learned weight updates while retaining the roughly 28-billion-parameter backbone. It reduces generated work while leaving the base model’s weight footprint largely unchanged.

Training the model to stop looping

UkisAI says its training process identifies reasoning-marker tokens associated with repeated deliberation, then penalizes those tokens during fine-tuning. Phrases such as “wait,” “actually,” and “let me reconsider” can precede another pass through the same reasoning. Reducing those patterns encourages the model to finish sooner, and the team reports fewer errors caused by excessive reconsideration.

The recipe also incorporates a transfer component from BottleCap AI’s ThinkingCap-Qwen3.6-27B, although the published results do not isolate its contribution. Training ran on eight NVIDIA H100 GPUs through NVIDIA’s Innovation Lab.

Math pays the largest price

The reported evaluation compares the BF16 Qwen3.8-27B base model with the same model plus the Swift adapter. BF16 refers to the 16-bit numerical format used for the model weights. The benchmarks cover general reasoning, mathematics, coding, terminal agents, and multimodal question answering. All figures are vendor-reported, with no confidence intervals or run-level variance included.