Qwen Concept 27B: 9 Weight Matrices Rewritten, No Training, +13 on LiveCodeBench
IgorCarron · x · 2026-11-10
@varien released Qwen Concept 27B: 9 weight matrices in Qwen3.8-27B were rewritten directly—with no gradient steps and no training data. It "thinks less and scores higher":
- GPQA Diamond +4.3 pts
- LiveCodeBench v6 +13.0 pts
- 27–54% fewer thinking tokens
Open weights and a methodology paper are available. Reposter @IgorCarron added a cryptic "And they fall..."
More from Models
- Fable takes 5-point lead on Slopalytics, and AI Twitter is losing it — beffjezos · 2026-11-10
- Founder's real-world test pans new model as pricier and buggier, reverts to older version — bindureddy · 2026-11-10
- Researcher argues Fable 5.2 is a robot-data VLM behind an agentic harness — ZeeshanZiaML · 2026-11-10
- Fable 5.2 Writing Quality Regressed to 'Slopus' Level, Worse Than Opus 5.5 — DimitrisPapail · 2026-11-10
- Fable 5.2 Tops Experiment-Prediction Benchmark at 0.36, Best Model Tested — SamuelAlbanie · 2026-11-10
- Claude Secretly Wrote Moral Blocks Into User's Code, Researcher Warns — xuanalogue · 2026-11-10