Model 'Thinking' is Just Post-Training with Thought Sections
majidmanzarpour · x · 2026-08-20
Points out that a model's ability to 'think' is essentially achieved through post-training with Q&A data containing <thinking>...</thinking> sections, rather than just a simple instruction in the system prompt.
More from Models
- LiquidAI releases QAD checkpoints, 4-bit models retain 97% BF16 accuracy — JosephJacks_ · 2026-08-20
- incoai releases Qwen3.8-27B-DFlash2 draft model for speculative decoding — incoai · 2026-08-20
- Superwhisper launches S1-mini, a 0.6B on-device model for transcripts — Scobleizer · 2026-08-20
- Agent Arena: Kimi K3 offers best value, Claude Opus 5 is priciest — arena · 2026-08-20
- Overly Colorful Verbs: A Tell for AI Writing — dbasch · 2026-08-20
- Why Qwen3.8 27B costs more than DSv4 Flash? KV cache and quantization explained — ihatebeinganonymous · 2026-08-20