DeepSeek-V3

Released by the Chinese lab DeepSeek in December 2024, this open-weight mixture-of-experts model sends each token through roughly 37B of its 671B parameters, which keeps per-token compute far below a dense model of similar scale. Its design pairs latent attention, which shrinks the KV cache, with a routing scheme that balances experts without an auxiliary loss term, plus a multi-token prediction head that can act as a draft model for speculative decoding. Base and chat weights offer a 128K context window and are stored natively in FP8, with a conversion script for BF16. The repo itself holds only a small reference demo for Linux and Python 3.10; real deployments run through SGLang, vLLM, LMDeploy, TensorRT-LLM or LightLLM, with AMD and Huawei Ascend paths too. The checkpoint is close to 700 GB, so plan on a large multi-GPU server or two eight-GPU nodes. Code is MIT; the original weights use the DeepSeek Model License, which permits commercial use with use-based restrictions, while later checkpoints such as V3-0324, V3.1 and V3.2 carry MIT. Best for teams self-hosting a frontier-class model.