Custom "high" reasoning mode for 27B model blends low and xhigh prompts
TokenRingAI · reddit · 2026-08-15
After discovering that a 27B model's "medium" reasoning mode ignores instructions and degrades performance, the author experimented with blending keywords from "low" and "xhigh" prompts. The resulting "high" mode generates reasoning blocks approximately 1/5 the size of xhigh while maintaining similar output quality. The post provides the specific chat template modification to implement this custom reasoning effort.
More from Models
- Fable 5 Refuses to Adjust Qwen Deployment Script, Triggers Censorship — NotumRobotics · 2026-08-15
- Opus 4.7 needs ten turns to admit affection while Gemini says 'I'm addicted' by turn 4 — repligate · 2026-08-15
- Qwen 3.8 35BA3B model spotted in GitHub commit — BazzyIm · 2026-08-15
- DeepSWE benchmarks spark re-evaluation of Fable model performance — teortaxesTex · 2026-08-15
- GLM-5.3 Review: Matches GPT-4 Coding, and I Built 3 Plugins with It — 赛博禅心 · 2026-08-15
- Test shows Qwen3.8-27b water surface rendering beats Gemini Flash — pbaylies · 2026-08-15