Three model names including deepseek-v5 have appeared in DeepSeek Harness test files. Is V5 coming?

Three models that DeepSeek had never previously made public have appeared in the official GitHub repository of DeepSeek Harness: deepseek-v5, deepseek-v5-open, and deepseek-v5-next.

After a user logs into an account, the client requests a model list from the server's /models endpoint, and then uses that list to decide which models can be used and which require updating DSH itself to a new version first.

deepseek-v5 is the most basic model. It declares no client version requirement, and the built-in catalog of the current version does not recognize it, so it is hidden, and calling it reports MODEL_UNSUPPORTED.

deepseek-v5-open means it is open to all users. The declared minimum client version is 0.0.1, meaning any version of the client can use it directly, and it will appear normally in the model list.

deepseek-v5-next represents models for subsequent updates, which require upgrading DSH before use. The required minimum version is 99.0.0-alpha.1, while the currently public DSH version is still at 0.2.x.

Obviously, DSH will never update to version 99.0.0, so deepseek-v5-next is just a "never reachable" version casually written in tests to reliably trigger the "version too low" branch.

In other words, DeepSeek has written into the tests in advance the three situations that may be encountered when V5 is released: a V5 that can be used without updating DSH, a V5 that requires updating DSH first, and a V5 that the current version does not recognize at all.

Some netizens believe that the models in the tests are all fake data and that these are merely placeholders.

Although placeholders are indeed very common in the software industry, used to pave the way for future updates.

But the essence of a placeholder is for engineers to gate things in advance. When V5 actually launches, no matter which version of DSH a user has installed, the software will not have problems.

This guarantee is only meaningful under the premise that "V5 will indeed come." No one would design a staged rollout mechanism for a name that is not on the roadmap.

Moreover, there are so many choices for placeholder names that any string would do. Since DeepSeek chose V5, it means either V5 is already in training, or V5 is already in the future schedule.

Looking at this mechanism itself, you will find that what DeepSeek wants to solve is a "release efficiency" problem.

Previously, when a new model was to go online, the client also had to release a version in sync, and users had to update the software to the new version before they could see it. Now, the server only needs to change the content returned by /models, and the new model can be opened directly to users, which is the approach corresponding to deepseek-v5-open.

The repository's update records also include a change stating that "the DeepSeek API key adapter now updates the model list in the same way as account login," indicating that this mechanism is being rolled out to more entry points.

This means that when DeepSeek V5 is released, in addition to DSH, other software capable of calling DeepSeek models can switch to DeepSeek V5 instantly, without requiring users to update the client again.

What deepseek-v5-next corresponds to is more like preparation for "new architecture, new capabilities." If V5 requires new client capabilities to be usable, such as new inference parameters or new interaction methods, this path will come in handy.

With both paths prepared at the same time, it shows that DeepSeek has already thought very clearly about the specific launch approach for V5.

However, one more point should be added: these three V5 models were not uploaded today. The actual upload time was September 24, 2026, and the uploader was Jiang Yichen, a graduate of Tsinghua University.

From the GitHub repository page, Jiang Yichen is one of the core contributors to DSH. In the early days of DSH, his number of commits ranked second in the project, second only to Cui Tianyi, and he participated in a large amount of R&D work such as model adaptation, Web, and connection recovery.