DeepSeek, a Chinese AI company, has launched a preview of its latest model, DeepSeek-V4, which includes two versions: Expert with 1.6 trillion parameters and Instant with 284 billion parameters. The Expert model is designed for advanced tasks, boasting 49 billion active parameters, while the Instant version has 13 billion active parameters. Notably, DeepSeek-V4 is an open-weights model, meaning users can download it from Hugging Face and run it on their own hardware, although substantial hardware is required for optimal performance. The model is also equipped with a 1 million token context window, enhancing its capabilities in handling complex queries. According to DeepSeek’s testing, the Expert version outperforms many competitors in reasoning tasks, particularly in math and coding, and rivals closed-source models in world knowledge. For those unable to run the model locally, DeepSeek offers an API for access. Pricing details for the API are also available.