Logit penalty on Qwen models improves math accuracy

A user tested logit-bias penalties on words like "wait", "maybe" and "perhaps" across quantizations of Qwen3.5-4B in llama.cpp, running 50 random MATH-500 questions, building on a Meta paper on overthinking markers. The paper had not covered llama.cpp's quantization options.

• Qwen 3.8 Flash runs at ~21 tok/s on RTX 30602026-10-10
• Benchmark: ThinkingCap vs Swift vs Qwen 3.8-27B2026-09-25
• Balance of Power in Open Models: Chinese Firms Lead Open-Weight2026-09-23
• Yandex Releases 80B AliceAI Foundation Model2026-09-22
• Qwen 27B runs autonomously for 3 weeks on one 30902026-09-21