The DeepSeek v3 paper (and [model card](https://github.com/deepseek-ai/DeepSeek-V3/blob/main/README.md)) are out, after yesterday's mysterious release of [the undocumented model weights](https://simonwillison.net/2024/Dec/25/deepseek-v3/). Plenty of interesting details in here. The model pre-trained on 14.8 trillion …