DeepSeek is recruiting beta testers for Harness — and V4-Flash's agentic scores may have been run on it

This is the best-sourced item in this batch, and it is the only one that traces to a named person at the lab rather than to an anonymous card.

On August 2, DeepSeek's Harness team lead Cui Tianyi publicly posted a call for applications to an internal beta of an open-source Agent Harness project. Applicants are asked for a GitHub ID and a record of contributions to relevant open-source projects. Successful candidates sign a confidentiality agreement. Screenshots of the invitations are circulating in developer communities.

Around that, the claim: DeepSeek Code is the product, powered by the new Harness framework. Built for autonomous software engineering with planning, tool use and code execution. Long-running agent workflows with memory and repository awareness. The recent DeepSeek V4-Flash benchmark numbers were reportedly produced with Harness. Positioned directly against OpenAI Codex and Claude Code. Closed beta expected soon.

The detail worth pulling out is the benchmark one. If V4-Flash's agentic scores were produced by the model running inside Harness, then those numbers describe a system, not a model — and reproducing them with the open weights alone would not be possible, because the harness is the missing half. That is the same trap we flagged this week on MiniMax-H3, where the open weights and the hosted product are different systems. It is worth knowing before you benchmark against a published figure.

Note also what this is not. Harness is a framework and a beta programme, not a new model. DeepSeek's V4 line is what it would run on.

Grade: reported. A named DeepSeek employee recruiting in public is meaningfully stronger evidence than an anonymous leak card — but the product name, the capability list and the competitive positioning are still second-hand, and DeepSeek has published no docs page, no changelog entry and no signup for any of it.