deepseek-ai/deepseek-v3

DeepSeek-V3 is an open-source large language model, the kind of AI that powers chat assistants. The repository contains the model's description, technical paper, instructions for downloading the weights, evaluation results, and code for running the model locally. Its main appeal, per the README, is that it is a very large model whose performance is comparable to leading closed-source models, yet it was trained at a much lower cost and is released under an open license. Under the hood, DeepSeek-V3 uses a Mixture-of-Experts design (often shortened to MoE). The model has 671 billion total parameters, but for any single piece of input only 37 billion are actually used, which keeps the cost of generating each answer low while still letting the overall model be very capable. The README explains that the architecture builds on a previous version (DeepSeek-V2) and adds an auxiliary-loss-free load-balancing strategy and a Multi-Token Prediction training objective for better performance. The training was done in FP8 mixed precision, a numerical format that uses less memory, on 14.8 trillion tokens of text. After pre-training, the model went through supervised fine-tuning and reinforcement learning, including distilling reasoning ability from DeepSeek's R1 reasoning model. The context window is 128K tokens. You would use DeepSeek-V3 if you want a strong open-weights model to run locally or integrate into your own product, if you are doing research on MoE architectures, or if you want to evaluate a state-of-the-art model without relying on a commercial API. The code is released under an MIT license, while the weights are under a separate model agreement, with downloads on Hugging Face.