DeepSeek’s in-house Harness tool is reportedly set to begin closed beta testing later this week, based on recruitment screenshots circulating in online communities. The images indicate that the first batch of users will be drawn from a small, limited group, and selected participants must submit personal information and sign a confidentiality pledge before they can gain access.
The claim was first circulated by Max For AI and later compiled by BlockTempo. According to the screenshots, a leak would carry consequences beyond the immediate program: the participant would lose current access and could also face reduced chances of being selected for future DeepSeek model tests, product betas, and cooperation opportunities.
Leak penalties reportedly extend beyond a single test
That wording matters because it pushes the cost of disclosure beyond a one-time removal. Under the terms shown in the screenshots, the impact would stretch across future dealings with DeepSeek, not just this round of testing.
For a company that in April uploaded V4-Pro weights to Hugging Face and open-sourced them under the MIT license, the shift has drawn attention. The report frames it as a distinction between model access and product-layer access: the model may be released, while the engineering layer around it remains tightly controlled.
BlockTempo’s report notes that model value sits in the weights, while a tool like Harness derives much of its value from product experience and how users actually interact with it. Open-sourcing that layer would expose product design directly. Anthropic’s Claude Code, the product Harness is said to be targeting, is also closed-source.
Hiring for the project appeared in May
The Harness effort did not surface only this week. In mid-to-late May, DeepSeek senior researcher Chen Deli posted hiring notices on X and Xiaohongshu for two Beijing-based roles: “Agent Harness Product Manager” and “Agent Harness R&D Engineer.”
The job description was direct. It said that everything outside the model itself falls within the scope of Harness, and summarized the idea with the phrase “Model + Harness = Agent.” The recruitment post also included the line “resumes go directly to the department head” and drew more than 300,000 views.
The reported team lead joined from Jane Street
According to a May report cited by BlockTempo, former TSY Capital co-founder Cui Tianyi joined DeepSeek in March to lead the newly formed Harness team and oversee a desktop Agent product aimed at Claude Code.
The same report said Cui spent nine years at Jane Street. It also described him as having won ACM Asia regional gold medals six times during university and as the author of the widely circulated Chinese programming text Nine Lectures on the Knapsack Problem.
BlockTempo linked that background to the nature of the product. In that framing, quantitative trading depends not only on strategy but on whether the execution system can turn strategy into results. Agent products face a similar test: model strength matters, but the outer harness is what turns intelligence into a usable product.
DeepSeek has not released a formal V4 version yet
The leak also suggested that if the information is accurate, the formal release of V4 and Harness could arrive soon. That possibility rests on one clear point: a formal V4 release has still not been published.
DeepSeek’s official API documentation shows its latest changelog entry as “DeepSeek V4 Preview Release” on April 24. At that time, the company open-sourced V4-Pro and V4-Flash. V4-Pro was listed with 1.6 trillion total parameters and 49 billion active parameters, while V4-Flash was listed with 284 billion total parameters and 13 billion active parameters. Both were released with a default context length of 1 million tokens. Three months later, the changelog has not been updated, and it still contains no formal announcement related to Harness.
Chinese internet posts have circulated claims such as a full commercial launch of the formal V4 release in mid-July and a time-of-day pricing model that would cut nighttime compute fees by 60%. BlockTempo said it checked DeepSeek’s official pricing page and found only two listed models, deepseek-v4-flash and deepseek-v4-pro, alongside a single pricing table, with no reference to time-based pricing.
Old API aliases were retired on July 24
One official change that did occur is the retirement of the old API aliases deepseek-chat and deepseek-reasoner at 15:59 UTC on July 24. BlockTempo noted that the change happened four days before this report.
For now, the Harness beta claim has been raised only by Max For AI. BlockTempo said it could not independently verify the authenticity of the screenshots, and DeepSeek had not responded at the time of publication. Based on the timeline shown in the images, a clearer answer may come later this week.
What is known so far
- DeepSeek Harness is said to enter closed beta later this week.
- The first group of users would be selected from a small pool.
- Selected testers would need to submit personal information and sign a confidentiality pledge.
- The screenshots say leaks could affect future eligibility for DeepSeek tests, products, and cooperation opportunities.
- DeepSeek had already started hiring for Harness-related roles in Beijing in May.
- The company’s latest official API changelog still stops at the April 24 V4 Preview release.
- DeepSeek has not publicly confirmed the beta or responded to the report.