Overview
- DeepSeek made the DeepSeek‑V4‑Pro‑0813 model and its weights public under an MIT license and published a developer preview of DeepSeek Harness v0.1 with a companion plugin ecosystem.
- The model is described as a mixture‑of‑experts with about 1.6 trillion total parameters, roughly 49 billion parameters activated per token, a claimed 1,000,000‑token context window and support for very large outputs and structured JSON and Tool Call APIs.
- Harness uses a Cordis 'everything‑is‑a‑plugin' design so models, tools, skills, sessions, sandbox, storage, loops, scheduling and UI are interchangeable plugins that developers can replace or extend without changing core code.
- DeepSeek published internal benchmarks reporting large gains over prior previews and scores that approach or exceed Claude Fable 5 on some tests, but those performance claims have not been independently validated.
- The MIT weight release and open plugin stack aim to let enterprises and developers self‑host on general GPUs and avoid vendor lock‑in, though operating a large MoE still requires significant compute and engineering effort and early community adoption is just beginning.