Overview
- DeepSeek published the DeepSeek-V4-Pro-0813 model weights and released DeepSeek Harness v0.1 as MIT-licensed projects, and the repositories are now publicly accessible so developers can download and run the code locally.
- Harness is built on the Cordis meta-framework and uses an “everything-is-a-plugin” design so models, tools, skills, storage, UI and other agent capabilities can be added, swapped, or extended without changing the harness source.
- DeepSeek describes V4 Pro as a mixture-of-experts model with 1.6 trillion total parameters, about 490 billion parameters active per token, a 1,000,000-token context window and support for structured JSON output and tool calls, but those technical claims come from DeepSeek’s documentation.
- The company published internal benchmark results that it says show large gains over preview builds and scores that approach or exceed Anthropic’s Claude Fable 5 on select tests, and those comparisons have not been independently verified.
- DeepSeek frames the open-source release as a way to let enterprises run models on-prem or on domestic and general-purpose GPUs, lowering vendor lock-in, though MoE models still demand substantial compute and engineering to run efficiently.