Swift-Qwen3.8-27B Cuts Reasoning Tokens 40%: Aider Benchmarks Show 2x Speed at Same Accuracy
returnity · reddit · 2026-09-16
UkisAI released Swift-Qwen3.8-27B, targeting Qwen3.8-27B's overthinking: the team identified reasoning-marker tokens that trigger overthinking in Qwen's reasoning rollouts, penalized them with RL, and added a transfer component derived from BottleCap AI's ThinkingCap-Qwen3.6-27B (a fine-tune already proven to save 40% tokens at comparable quality).
The poster independently verified the claims with an Aider-based coding eval suite: Pass1 30.8% vs 27.1%, Pass2 75.7% vs 77.6% — essentially parity — while completion tokens dropped from 12,547 to 7,301 (63%) and seconds per case halved from 1,481 to 750. Because decode speed degrades with longer responses, token savings translate into outsized time savings. A big win for local 27B users.
More from Models
- Claude upgrade-page traffic rebounded to 17.6M in September, near May's 18.6M peak — firstadopter · 2026-11-02
- Open-source GLiNER-PII scrubs 13 identifiers from a discharge note in 550ms on GPU — MaziyarPanahi · 2026-11-02
- OpenAI Dots user says bloat and broken infra drove him to try newcomer Muse — Koala_Confused · 2026-11-02
- Fable 5.5 reportedly one-shots animation in pure code as Anthropic's coding bet pays off — Hesamation · 2026-11-02
- One agent goal has been running for 4 days, burning 30K tokens — CtrlAltDwayne · 2026-11-02
- Daily AI Brief: Gemini 4 Carbon spotted on Arena, DeepSeek open-sources coding agent — testingcatalog · 2026-11-02