DeepSeek, a China-based AI startup, said on July 31 that its DeepSeek-V4-Flash API was in public beta.

The move broadened developer access after preview releases in April and came ahead of the official release of DeepSeek-V4-Pro, which the company said “will follow soon.”

In earlier documentation, DeepSeek described V4 Flash as a faster, lower-cost model with 284 billion total parameters and 13 billion active parameters. The model supports 1 million tokens and offers thinking and non-thinking modes.

The company said the model was built for coding and tool-driven agent workflows. Developers can use existing SDKs compatible with OpenAI- and Anthropic-style APIs, as well as coding tools such as GitHub Copilot and OpenCode.

DeepSeek said V4 Flash had “significantly enhanced agent capabilities” and benchmark results “far exceeding V4-Pro-Preview.”

However, the materials reviewed did not include a detailed public score table. Instead, the July 31 changelog described testing for code-agent tasks and referenced internal DSBench evaluations.

🔗 Source: Bloomberg