DeepSeek-R1

DeepSeek-R1 launched 20 January 2025 as the reasoning splash: large-scale RL, MIT license, distilled smaller models, DeepThink on chat. The May 2025 R1-0528 note added benchmark gains, fewer hallucinations, JSON, and function calling. The official footer still points “DeepSeek R1” at that 0528 note, so I keep R1 in the lineup as a research chapter.

Today’s Thinking Mode is the descendant of that DeepThink idea, now living on V4 with effort levels. I do not call R1 from the current pricing table unless a self-hosted weight is the point of the experiment.

If you want the current hosted reasoning path, use Thinking Mode on V4. If you want the 671B MoE that came just before R1’s fame, see DeepSeek-V3.