Skip to content
← All news
5 min read

DeepSeek shipped V4 Pro and an MIT-licensed answer to Claude Code

The 1.6 trillion parameter flagship left preview on August 13, next to Harness v0.1, a plugin-based agent framework under MIT.

A 1.6T flagship out of preview, agent scores tripled, and an MIT agent framework at 23,000 stars in hours.

DeepSeek moved its 1.6 trillion parameter flagship out of preview on August 13. V4 Pro, build 0813, closes a testing window that opened in April, and it arrived with a second release that may matter more: DeepSeek Harness v0.1, an agent framework published under the MIT license.

The benchmark jumps DeepSeek reports

  • Terminal-Bench 2.1: 72.1 to 87.9.
  • CyberGym: 52.7 to 83.3.
  • DeepSWE: 12.8 to 62.7, roughly a 5x jump on the agentic software engineering benchmark.

Those are DeepSeek's own numbers, so apply the usual discount until independent runs land. What is not self-reported: VentureBeat notes the GA release came with higher API prices, and the Harness repository passed 23,000 GitHub stars within hours of publishing.

Harness is the interesting half. It treats everything as a plugin: models, tools, skills, and sessions can be replaced and recombined, with four operating modes out of the box. VentureBeat calls it an open-source rival to Claude Code, and the MIT license means teams can embed it in commercial products without a per-seat meter.

Agent harnesses are where lock-in actually forms. Models swap in minutes, but workflows, skills, and session formats calcify. An MIT-licensed harness is a bid to own that layer in the open.

Why a build studio cares

We ship agent workflows for clients, and the harness question comes up in every scoping call: what runs the loop, and what does it cost at 50 seats? A credible open harness changes that math. The tradeoff is that v0.1 software days old is v0.1 software: it needs a real security review before it touches client data, especially a framework whose whole design is loading third-party plugins.

Next step: run Harness in a sandbox against a throwaway repo, give it the same task you last gave your current agent tool, and diff the transcripts. An hour of that beats any benchmark table, including the one above. If you are choosing an agent harness for a client deployment, write to us at hello@gattyworks.com.

DeepSeekAI AgentsOpen SourceCoding ModelsDeepSeekDeepSeekHarnessV4ProAIAgentsOpenSourceMITLicenseClaudeCodeCodingModelsAgentFrameworksAINews

Ready to know?

Send what you want checked or built. Fixed scope, price, and date in writing inside 24 hours, or the website or audit fee on your first project is refunded in full.

24 clock hours. Weekends included.