DeepSeek Harness Release Date: v0.1 Developer Preview Is Here

DeepSeek Harness was released as a developer preview on August 13, 2026, version v0.1, with source code open-sourced under the MIT license the same day^[1][2]. It is live today: the official page at deepseek.com/harness is up, the GitHub repository is public, and the npm package @deepseek-ai/dsh installs and runs^[1].

If you are asking 'when is DeepSeek Harness coming out?' because you saw 'to be released soon' in the July 31 changelog — that note is out of date. The framework has shipped, and the current version is a developer preview rather than a stable 1.0^[1][2].

The name 'DeepSeek Harness' first appeared in official documentation in the July 31, 2026 changelog that launched the V4-Flash-0731 build^[4].

The changelog stated that every Code Agent benchmark for the new build was run using 'the DeepSeek Harness minimal mode (to be released soon)' with max effort, topp=0.95, and temperature=1.0. The Hugging Face model card said the same. That single mention is what created the 'to be released soon' expectation that dominated the first half of August^[4].

• July 31, 2026: changelog names DeepSeek Harness minimal mode as the benchmark framework^[4]
• Same day: media and community start tracking 'when does DeepSeek Harness release'
• No repo existed under github.com/deepseek-ai at that point

Before the launch, two leaked-report threads pointed at a mid-August window. On July 28, Fast Technology (via Sina Finance) reported the official V4 rollout would begin with internal testing around mid-August. On July 30, Zhineng Jiyuan AGI (via Sina) reported an NDA-style closed beta for a small group of invited users, followed by general availability between August 10 and 20^[5].

Both reports relied on circulated screenshots and anonymous sources, so they were never independent confirmation. They turned out to be directionally correct: the public release landed on August 13, inside the reported window^[5][1].

On August 13, 2026, DeepSeek announced the developer preview on its official channel: 'DeepSeek Harness is now in developer preview for agent harness developers worldwide — source code included'^[1][6].

The launch bundle was: the official page (deepseek.com/harness/en), the GitHub repository (deepseek-ai/deepseek-harness), the npm package @deepseek-ai/dsh, developer docs, a community plugin index, and the Cordis paper. Everything is a plugin became the tagline: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI are all swappable plugins^[1].

The market response was immediate. Star counts went roughly 38K on day one, ~95K by day two, and over 127K with 12.7K+ forks by August 16. An independent plugin index already tracks 616+ community plugins^[7][8].

• Aug 13: official announcement on X (@deepseek_ai) and the harness site^[6]
• Same day: source on GitHub, package on npm, docs live^[1]
• Aug 14–16: 127K+ stars, 616+ plugins indexed, multiple independent guides appear^[7][8]

A developer preview is not a stable release. The official README states plainly: 'THERE WILL BE COMPATIBILITY-BREAKING CHANGES'^[3].

In practice this means: the core architecture (Cordis plugin kernel) is the point of stability, while plugin APIs, CLI flags, and configuration formats can change between releases. The npm package currently ships 0.1.0-rc builds. Projects that build on dsh today should pin versions and follow GitHub Discussions for breaking-change announcements — the project deliberately does not use GitHub Issues^[3].

The likely next milestones, based on the official roadmap signals^[1][3]:

For hands-on steps today, jump to How to Install DeepSeek Harness. For the release that shipped alongside it, read DeepSeek V4 Pro GA release.

• A stable 1.0 release (no date announced) that freezes plugin APIs
• More official plugins in the models/tools/sandbox categories
• DeepSeek's own agent product built on the harness, hinted at by the hiring materials
• Third-party reproductions of the official benchmarks using minimal mode (see harness benchmarks)