DeepSeek releases official V4-Flash model, 284 billion parameter open weight model for agents
DeepSeek released the official V4-Flash model on July 31: a 284-billion-parameter open-weight model with a 1-million-token context window, built for coding and agents. It outscores DeepSeek's own larger V4-Pro-Preview on all nine published agent and coding benchmarks.
DeepSeek officially released DeepSeek-V4-Flash-0731 on July 31, the production version of its Flash model following an April preview. It is a 284-billion-parameter Mixture-of-Experts model with 13 billion active parameters and a 1-million-token context window, built for coding, tool use and agentic workflows. The company said the model's architecture is unchanged from the preview, with all performance gains coming from extensive post-training. The 0731 build scores higher than DeepSeek's own V4-Pro-Preview on all nine agent and coding benchmarks the company has published. The API is now in public beta, and the model can also be downloaded and run on private servers. That a smaller, cheaper model outperforms DeepSeek's own larger flagship on every benchmark the company has released challenges the assumption that bigger open models are automatically the better choice for agent workloads, and gives developers a cheaper option as they build autonomous AI agents.