Release anniversaries, 26 Dec 2026
DeepSeek-V3 release
On 26 December 2024 DeepSeek released V3, a 671-billion-parameter open mixture-of-experts model.
The Hangzhou lab DeepSeek released DeepSeek-V3 on 26 December 2024 with open weights. It is a mixture-of-experts model with 671 billion parameters in total, of which about 37 billion are used for each token, trained on 14.8 trillion tokens. Its technical report said the full training used about 2.8 million hours on Nvidia H800 chips, a small fraction of what rival frontier models were believed to need, helped by low-precision FP8 training and its multi-head latent attention design. It was the base for DeepSeek-R1 a month later.