DeepSeek Unveils V4 Pro and V4 Flash AI Models

Chinese artificial intelligence startup DeepSeek unveiled preview versions of its latest large language models, DeepSeek-V4 Pro and DeepSeek-V4 Flash, on Friday, positioning them as competitive challengers to leading systems from U.S. companies like OpenAI and Google. Based in Hangzhou, the firm released the open-source models to developers for testing and feedback, highlighting their advanced capabilities in reasoning, coding, and processing extensive data.

The V4 Pro model, with 1.6 trillion parameters, outperforms other open-source models in mathematics, coding tasks, and world knowledge benchmarks, according to DeepSeek. The company stated that it trails only Google's closed-source Gemini 3.1 Pro in general knowledge assessments and performs comparably to top proprietary systems, lagging by an estimated three to six months in development. DeepSeek described the model as establishing itself as the leading open-source option available. Its companion, the V4 Flash, features 284 billion parameters and offers similar reasoning power with faster response times and lower operational costs, making it suitable for broader applications.

Both versions support a context length of one...