DeepSeek Harness is rumored to have begun internal testing this week, going head-to-head with Claude Code in capability

DeepSeek’s own Harness tool has reportedly leaked to start its internal testing later this week, and a recruitment screenshot shared by the community shows that the first batch of users will be selected from a small set of groups. Those selected must submit personal information and sign a “Non-Disclosure Commitment Letter” to receive access to the product. The screenshot also states that if any data leak occurs, it won’t only result in loss of that round of eligibility, but will also affect future selection for all kinds of DeepSeek models, product internal tests, and partnership opportunities. According to a tip from Max For AI, compiled by 動區動趨.
(Background: DeepSeek V4 is released—can its programming capabilities beat GPT and Claude? Costs again topples the charts)
(Background addition: OpenAI scientists recommend: don’t put too much effort into Harness; the next-generation model may be built-in)

Table of contents

Toggle

• The cost of leaking
• The leader has spent nine years at Jane Street
• The V4 full version still hasn’t arrived yet

Key takeaways

• Max For AI claims that DeepSeek Harness will start internal testing later this week, and the first batch of users will be selected from a small group
• Selected users must sign a “Non-Disclosure Commitment Letter”; leaking will also affect eligibility for all future DeepSeek products and partnerships
• The official update log still stops at the V4 Preview from April 24; the full version and Harness have not been published

Three months ago, on DeepSeek’s official release page for the V4 Preview, it was still teaching developers how to connect the model into Claude Code. Three months later, what it’s preparing to open for internal testing is its own product built specifically to work with Claude Code.

On July 28, Max For AI reported that DeepSeek’s internally developed Harness tool is about to begin internal testing. The recruitment screenshot shared by the community shows that the plan is to start later this week. The first batch of users will be screened from a small set of groups; those selected must submit personal information, sign a “Non-Disclosure Commitment Letter,” and only then will they receive access to the product.

The cost of leaking

According to the screenshot, once there is a leak, not only will the eligibility for that round be canceled, but it will also affect selection for subsequent DeepSeek models, product internal tests, and partnership opportunities.

The weight of this line is that it expands the cost of breach from a one-time penalty to impacting all future interactions. The people who sign are not putting collateral on the test eligibility this time—they are staking their position within the DeepSeek ecosystem.

For a company that only launched the 1.6 trillion-parameter V4-Pro weights directly onto Hugging Face in April and open-sourced them under the MIT license, this shift itself is a signal. Models can be shipped, but the engineering layer must be locked down.

The logic is not hard to understand. The value of the model is in the weights. Once the weights are open-sourced, others still need to prepare their own compute to run it. Harness’s value lies in the product experience and how users actually use it—opening something like this is like handing the product directly to a rival. Anthropic’s Claude Code is also closed source.

The one leading the team has spent nine years at Jane Street

DeepSeek’s Harness team was not formed only this week. In mid-to-late May, DeepSeek senior researcher Chen Deli posted recruitment information on X and Xiaohongshu. It was hiring for two roles: “Agent Harness Product Manager” and “Agent Harness R&D Engineer,” with the work location limited to Beijing. The job descriptions were written very plainly: “Except for the model itself, all other work falls within the scope of Harness.” The core formula is “Model + Harness = Agent.” That post also included a line about “a resume going straight to the department head,” and it drew more than 300k views.

The choice of leader is even more worth looking at. According to a May report, former TSY Capital co-founder Cui Tianyi joined DeepSeek in March this year and became the head of the newly formed Harness team, leading a desktop agent product that is designed to match Claude Code. He spent nine years at the Wall Street quant giant Jane Street. During college, he won the ACM Asia regional gold medal six times, and he wrote “Nine Lectures on the Knapsack Problem,” a text that has circulated for years in Chinese programming circles.

Readers of 動區 should be familiar with the name Jane Street—that’s the firm with $39.6 billion in revenue last year and average compensation of $2.68 million per person. DeepSeek hiring people with quant backgrounds to build programming tools makes sense in terms of logic. The core of quant trading has never been just having smarter strategies—it’s whether the execution system can turn strategies into money. The same applies to agents: no matter how strong the model is, what truly turns intelligence into a product is the harness layer outside it.

The V4 full version still hasn’t arrived yet

The other half of the leak is: “If the news is true, the V4 full version and Harness will be released soon.” For this statement to hold, the V4 full version must still not have been released—up to today.

And indeed, it hasn’t. In the latest entry of the DeepSeek API official documentation update log, the newest one is still “DeepSeek V4 Preview Release” dated April 24. The “Preview” is labeled by the official itself. At the same time they open-sourced V4-Pro (1.6 trillion total parameters, 49 billion activated) and V4-Flash (284 billion total parameters, 13 billion activated), with a default context length of 1 million tokens. Three months have passed; the update log hasn’t moved again, and there has been no entry related to Harness.

On Chinese internet, however, there are already plenty of claims like “full commercial launch of the V4 full version in mid-July” and “first-in-peak-and-valley time-of-use pricing, with night compute costs dropping by 60%.” 動區 checked the DeepSeek official pricing page: it only lists two models—deepseek-v4-flash and deepseek-v4-pro—and a single pricing table; there is no time-of-use pricing.

The only real official move that has happened is different. The old API aliases deepseek-chat and deepseek-reasoner were fully retired at 15:59 on July 24 (UTC), just four days ago. Clearing out old entry points usually makes room for the new versions.

Right now, this internal testing leak has been raised only by Max For AI. 動區 cannot independently verify whether the screenshot is genuine, and DeepSeek has not responded. Based on the schedule given in the screenshot, the answer will be known later this week.

Common questions

What is DeepSeek Harness?

DeepSeek’s internally developed agent tool targets Anthropic’s Claude Code. The team opened recruitment in May in Beijing; the core formula is “Model + Harness = Agent.” It is responsible for tool calling outside the model, context management, and terminal execution.

Has the DeepSeek V4 full version been released?

Not yet. The latest official API update log entry is still the V4 Preview from April 24, which open-sourced V4-Pro and V4-Flash. The online rumors about “full commercial use in July” and “peak-valley time-of-use pricing” cannot be found in the corresponding official pricing page.