Intel / article

Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh

Source: huggingface.co
Author: kisjovan
Date: Published Sep 16, 2026

Context

Swift-Qwen3.8-27B is a retrained version of Qwen's 27-billion-parameter model, built to reduce repetitive reasoning. Its developers say they identified reasoning-marker text units linked to overthinking and penalized them during additional training. Their model card reports that, on GPQA-Diamond at the highest reasoning setting, accuracy changed from 88.38% to 88.28% while median thinking output fell from 6,642 to 2,771 .

These are the developers' own benchmark results, and the tradeoff varies by task. On AIME 2026, the same comparison reduced accuracy from 98.67% to 94.00%. The supported finding is that targeted training removed a large share of reasoning output while keeping some benchmark results close to the base model; the headline's accuracy claim does not hold equally across the evaluation.

Key takeaways · AI-distilled

Swift-Qwen3.8-27B was post-trained by identifying and penalizing tokens linked to 'overthinking' rather than truncating reasoning length directly, then used on-policy (aided by Nvidia-donated GPUs) to recover accuracy lost from that penalty.
On GPQA-Diamond, Swift matched the base model's accuracy (88.3% vs 88.4%) while cutting median reasoning tokens 58%, and at xhigh effort it beat the base model's own medium-effort accuracy by 4 points while using about 1.6x medium's token count.
Token savings held across all reasoning-effort settings (mean thinking reduction of 41% at xhigh, 23% at medium, 26% at low), though the team notes medium and low effort saw 1-4% and 1-2% accuracy drops that need further testing.
The team says their method isn't a replacement for reasoning-effort settings or token caps — it targets a separate 'anxiety-like' reasoning-loop pattern they say is prominent even in BF16 models of this size class, not just ones.
The model's license is not Apache 2.0 — it only restricts companies with over $1M in revenue — which the team flagged as an open question for community feedback rather than a settled decision.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.

Read the source huggingface.co