Late on August 12, DeepSeek quietly brought V4 Pro to general availability under the version number DeepSeek-V4-Pro-0813. There was no WeChat post and no technical blog. The first signal anyone outside the company noticed was the version number on the API pricing page switching from preview to release. Less than 24 hours later, the launch banner on the website and the notice on the open platform both disappeared. DeepSeek has offered no public explanation.
But right after the V4 Pro launch materials vanished, another product the company had been preparing went ahead as planned: DeepSeek Harness entered Developer Preview and was open-sourced. One launch quietly contracting while an agent infrastructure project opens up on schedule — that contrast is the most notable thing about the past 24 hours.
What happened in 24 hours
The timeline is straightforward. Late on August 12, DeepSeek-V4-Pro-0813 went live. Through the day on August 13, Chinese tech media picked it up, corresponding launch materials appeared on the website and open platform, and the first users began testing. By later on August 13, the launch banner and the open-platform notice were both gone.
One distinction matters here: what can be confirmed is that the launch materials came down, not that the model left the product lineup. V4 Pro is still visible on the API pages and the model name is still there. Rather than calling this a delisting, it is more accurate to say DeepSeek has temporarily withdrawn the public signal that V4 Pro had formally launched.
As for why, the company has said nothing. Plenty of theories are circulating, but until DeepSeek explains, any specific cause remains speculation. What is worth watching is whether the company relaunches, revises the version, or addresses the release at all.
Where the dispute begins
The controversy around V4 Pro is not that the model does not work. It is that a visible gap opened between the capabilities shown in official figures and what some users experienced.
At launch, DeepSeek published a set of agentic benchmarks comparing V4 Pro against frontier models including Claude Fable 5. By those numbers, V4 Pro comes close to or surpasses its competitors on multiple agent tasks. But as the first wave of users started working with it, that lead did not fully reproduce.
The recurring feedback is that the real-world improvement over the Flash version is less pronounced than the official tests suggest, particularly in coding, tool calling and long-running tasks.
None of this proves the official tests were flawed. The more interesting question is what sits between a benchmark and real usage. For agents, that variable may not be the model alone.
What V4 Pro's scores were actually measuring
Start with the model. According to DeepSeek, V4 Pro uses a mixture-of-experts architecture with roughly 1.6 trillion total parameters, about 49 billion activated per token, a 1 million token context window and a maximum output length of 384,000 tokens. On pricing, V4 Pro continues DeepSeek's low-cost API positioning, sitting far below Claude Fable 5.
What makes V4 Pro interesting is not the parameter count or the price on its own, but DeepSeek's attempt to strike a new balance between capability, inference cost and long context. The company has pointed to a new generation of attention mechanisms and KV cache optimizations aimed at lowering the real cost of running a model this size.
The problem lies with agent testing. How much work an agent can actually finish does not depend only on whether the model can think. It depends on how it calls tools, manages context, handles errors, runs task loops, and interacts with the filesystem and runtime environment.
For agents, the honest formulation is not that the model is the agent. It is that model capability plus execution framework is what produces the agent capability a user sees.
That is what makes Harness central to understanding this dispute.
What Harness fills in
At the same moment the V4 Pro launch materials came down, DeepSeek Harness was open-sourced on schedule.
Per the official repository, Harness is in Developer Preview under an MIT license, built on a single design principle: everything is a plugin. Models, tools, skills, sessions, sandboxes, filesystems, task loops, orchestration and UI can all be composed and swapped as plugins. The project is built on the Cordis meta-framework, letting developers reshape agent capability through configuration rather than by editing the underlying source.
Put simply, the model handles reasoning and Harness turns that reasoning into action.
When a model needs to read files, call tools, edit code, run commands or work through a complex task over time, something has to manage that whole process. That layer is what Harness addresses.
It is also what separates Harness from an ordinary model SDK. An SDK answers the question of how to call a model. Harness answers the question of how to keep a model working until a task is done.
Why open-sourcing Harness matters
If part of the official agent testing for V4 Pro ran inside DeepSeek's own execution environment, then it is hardly surprising that outside users calling the model directly through their own toolchains would get different results.
This is not a refutation of the official numbers. It points at a newer problem in agent evaluation: are we comparing models, or comparing entire agent systems?
Older model evaluation could treat the model itself as the core variable. But as AI moves from answering questions to executing tasks, tool calling, context management, sandboxing and task loops matter more and more. The same model, dropped into different execution frameworks, can perform noticeably differently.
So open-sourcing Harness is not just one more open-source release. It offers a possibility: that outside developers can now run the published agent numbers again in a framework much closer to DeepSeek's own test environment.
The boundaries still hold, though. Open-sourcing Harness does not by itself validate the earlier official results, nor does it automatically explain the gap between user experience and official data. What it provides is a more public entry point for verification. The answer still has to come from third-party testing.
One signal contracts, one door opens
Put the two moves side by side and the past 24 hours read more clearly.
On one side, V4 Pro's launch materials vanish. On the other, Harness enters Developer Preview exactly as planned. The first leaves a question unanswered: why did DeepSeek adjust how V4 Pro was announced? The second hands a different question to developers: if there is a gap between official data and real experience, can it now be re-tested under conditions closer to the company's own?
Seen that way, the value of open-sourcing Harness lies in verification.
For DeepSeek, opening Harness means handing developers agent execution capability that had been internal. If third parties running the same framework land near the official numbers, V4 Pro's earlier data gains outside support. If a clear gap remains, the questions move on to whether the issue is model capability, execution framework or evaluation method.
That is worth more than publishing another set of benchmarks. A benchmark gives you a result. Opening the tooling gives more people the chance to check that result themselves.
Three things to watch
First, whether DeepSeek explains why the V4 Pro launch materials were pulled. The model remains visible and callable within the product lineup, but the public signal of a formal launch has changed. Whether the company relaunches, revises the version, or simply rewrites the notice is still unknown.
Second, whether third-party evaluators re-test V4 Pro using DeepSeek Harness. That will determine how far this dispute can be clarified. If results under the same Harness come close to the official figures, the earlier experience gap may have more to do with user toolchains. If a significant gap persists, the question returns to the model itself and to evaluation methodology.
Third, whether Harness builds a real developer ecosystem. DeepSeek has competed primarily on model capability and inference cost; Harness represents a push further into agent infrastructure. If the project ends up serving only DeepSeek's own models, its impact stays limited. If the plugin design attracts developers to assemble models, tools, sandboxes and workflows on top of it, then what DeepSeek is competing for is no longer just the model market but the agent infrastructure layer beneath it.
In closing
Over the past 24 hours, DeepSeek made two moves that point in opposite directions: V4 Pro's launch materials quietly disappeared, while Harness opened to developers on schedule.
The first suggests V4 Pro's formal release still has variables worth watching. The second suggests DeepSeek has not slowed its push into agent infrastructure because of it.
For now, rather than rushing to judge whether V4 Pro is any good, it is worth waiting for results that can be reproduced. Model performance is arguable. Benchmarks are arguable. But once Harness is in developers' hands, the act of verification itself becomes more open.
DeepSeek pulled back the signal announcing V4 Pro. It pushed out the tool for checking it.
The answer, from here, will not come only from DeepSeek. It will come from the people who actually get this system running.
Information in this article comes from DeepSeek's website and official GitHub repository, published reporting, and publicly posted user feedback. On model performance, benchmarks and user experience, readers should distinguish company self-testing from third-party testing; as of publication, some figures remain independently unverified. This article does not constitute investment or purchasing advice.