DeepSeek unveils V4-Pro-0813 open-source model and Harness agent framework

DeepSeek has open-sourced its V4-Pro-0813 model and launched the public beta of Harness, a coding agent framework, along with a plugin ecosystem. Harness turns models into agents by managing context, tools, and task states, aiming to complete coding tasks end-to-end. It competes with OpenAI's Codex and Anthropic's Claude Code.

The V4-Pro-0813 model uses a mixture-of-experts design with 1.6 trillion total parameters, activating only 49 billion per token. It supports a 1 million token context window and can output up to 384,000 tokens. It offers thinking and non-thinking modes, plus full support for JSON output, tool calls, and the Anthropic API. Internal tests show it scored 87.9 on Terminal Bench 2.1, up from 72.1 for the preview version.

The model weights are released under the MIT license, so developers can download and run them locally, reducing reliance on specific hardware and making it easier for governments and enterprises to deploy privately.