DeepSeek-V3

DeepSeek-V3 arrived 26 December 2024 as a 671B MoE with 37B active parameters, trained on 14.8T tokens, with a claim of 60 tokens per second — three times V2. The free app launched on this generation. The official research footer still lists it. I keep the page so the history does not collapse into a single V4 headline.

V3 is not what chat.deepseek.com served me on this visit. The banner and the API both pointed at V4. If you self-host, read the V3 card on Hugging Face for license and hardware notes. If you call the hosted API, start with DeepSeek-V4-Flash.

The jump from V3 to V3.1 was hybrid thinking and agents. The jump to V4 is the 1M context default and the Pro/Flash split. That is the plot. Everything else is a footnote I only keep because DeepSeek still publishes it.