A deep dive into how LLMs implement multiple reasoning effort modes (low, medium, high). Covers the foundations of reasoning models via RLVR, the role of think tokens, and how on/off reasoning switches work. Examines concrete implementations across six open-weight flagship models — DeepSeek V4, Nemotron 3 Ultra, Kimi K2.5, GLM-5, Qwen3, and Inkling — detailing their distinct approaches: separate effort specialists, mode-conditioned RL with token penalties, SFT data mixing, hard token budgets, and continuous effort conditioning. Also clarifies the distinction between training-scale and inference-scale knobs using GPT-5.6 as an example.
Table of contents
1. A brief definition of reasoning models2. A brief overview of training and inference scaling reasoning models3. Think tokens4. Reasoning mode on and off switches5. How “reasoning effort” settings work6. Bonus: Different ways to implement reasoning efforts (in flagship open-weight LLMs)6.5 Qwen3 uses mode fusion and inference-time truncation6.6 Inkling conditions RL on a continuous effort value6.7 Overview of the known recipes7. ConclusionFurther resources208 Impressions