DeepSeek-V3

DeepSeek-V3, released December 26, 2024, is a 671-billion-parameter mixture-of-experts language model with 37 billion active parameters per token. It was pre-trained on 14.8 trillion tokens — the first model at that scale trained primarily in FP8 precision — using multi-token prediction and an auxiliary-loss-free load-balancing strategy. DeepSeek reported a training cost of just $5.576 million (2.788 million H800 GPU-hours), roughly one-twentieth of comparable Western frontier runs, and benchmark results matching or beating GPT-4o and Claude 3.5 Sonnet, including on Codeforces and Aider Polyglot coding tests. Released under the MIT license with full weights, it immediately became the strongest open-weight model available and served as the architecture base for DeepSeek-R1 a month later.

Limitations: despite the MIT license, the model is of Chinese origin and organizations with strict data-sovereignty or compliance policies may need additional review. At 671B total parameters it requires substantial inference infrastructure (multi-GPU or API access); most users will run it via DeepSeek's API or a hosted provider. Later variants (V3.1, V3.2, V4) supersede it on reasoning and context length, and vendor-reported benchmarks should be treated as vendor-reported.