DeepSeek, a startup based in Hangzhou, China, released two new artificial intelligence models on April 24. The models, called V4-Pro and V4-Flash, are open-weights, which means anyone can download them and use them freely.
V4-Pro is the world's largest open-weight AI model, containing 1.6 trillion parameters. However, only 49 billion parameters work at any given moment, thanks to a special design called Mixture-of-Experts. This design splits the model into smaller networks and activates only the ones needed for each task.
The price is remarkable. V4-Flash costs just $0.14 per million input tokens through DeepSeek's API, while V4-Pro costs $3.48 per million output tokens. By comparison, OpenAI charges $30 and Anthropic charges $25 for similar services.
DeepSeek trained these new models using Huawei Ascend processors, Chinese-designed AI chips that offer an alternative to Nvidia's technology. The company says V4 performs almost as well as the world's leading models from OpenAI and Google, though it falls slightly behind them.
The models are text-only and do not process images, audio, or video. They are released under the MIT license, meaning developers can build commercial products using them without paying DeepSeek. This strategy represents DeepSeek's bet that the future of AI is free and accessible to everyone.