文 | 字母AI
DeepSeek Harness的官方GitHub代码仓库里出现了三个DeepSeek此前从未公开过的模型，分别为deepseek-v5、deepseek-v5-open、deepseek-v5-next。
这段代码来自于DSH里账号登录通路的测试文件provider.spec.ts。
用户登录账号后，客户端会向服务端的/models 接口要一份模型清单，再根据清单决定哪些模型能用，哪些要先把DSH本身更新到新版本。
deepseek-v5是最基础的模型，没有声明客户端版本要求，当前版本的内置目录不认识它，所以隐藏，调用时报MODEL_UNSUPPORTED。
deepseek-v5-open代表开放给所有用户。声明的最低客户端版本是0.0.1，也就是任何版本的客户端都能直接使用，会正常出现在模型列表里。
deepseek-v5-next代表着后续更新的模型，需要升级DSH才能使用。要求的最低版本是99.0.0-alpha.1，而目前公开的DSH版本还停留在0.2.x。
很明显，DSH永远不会更新到99.0.0版本，所以deepseek-v5-next只是测试里随手写的一个“永远达不到”的版本，用来稳定触发“版本不够”这个分支。
换句话说，DeepSeek把V5发布时可能遇到的三种局面，提前写进了测试：不用更新DSH就能用的V5，必须先更新DSH才能用的V5，以及当前版本压根不认识的V5。
一些网友认为，测试里的模型都是假数据，这不过是占位符而已。
虽然占位符在软件行业中确实很常见，用来给以后的更新铺路。
但占位符的本质是给工程师用来提前把关的。等V5真上线了，不管用户装的是哪个版本的DSH，软件都不会出问题。
这种保证只有在“V5确实会来”的前提下才有意义。没有人会为一个不在路线图上的名字，去设计分级发布的机制。
何况占位名有那么多选择，随便一个字符串都行，DeepSeek既然选的是V5，就代表要不然V5已经在训练了，要不然V5已经在未来的排期里了。
再看这套机制本身，你会发现DeepSeek想解决的是一个“发布效率”问题。
在以前，新模型要上线，客户端也得同步发版，用户得先把软件更新到新版才能看到。而现在，服务端只要改一下/models的返回内容，新模型就能直接对用户开放，即deepseek-v5-open所对应的方式。
仓库的更新记录里，也出现了“DeepSeek API key适配器现在和账号登录用同样的方式更新模型列表”的改动，说明这套机制正在被推广到更多入口。
这代表DeepSeek V5发布时，除了DSH以外，其他能调用DeepSeek模型的软件，都可以秒切DeepSeek V5，不需要再让用户更新一下客户端。
deepseek-v5-next对应的，则更像是为“新架构、新能力”准备的。如果V5需要新的客户端能力才能用，比如新的推理参数或者新的交互方式，这条路径就派得上用场了。
两条路径同时备好，说明DeepSeek对V5具体的上线方式，已经想得很清楚了。
不过还要补充一点，这三个V5模型并不是今天上传的。实际的上传时间是2026年9月24日，上传者是毕业于清华大学的蒋奕晨。
从GitHub仓库页来看，蒋奕晨是DSH的核心贡献者之一。DSH初期，他的提交数排名项目第二，仅次于崔添翼，参与了模型适配、Web、连接恢复等大量研发工作。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Letter AI
Three previously undisclosed models of Deepseek-v5, Deepseek-v5-open, Deepseek-v5-next, and Deepseek-v5-nextt were found in the official GitHub code warehouse of DeepSeek Harness.
The code is derived from the DSH account log-in test file provider.spec.ts.
After the user logs in, the client asks the service-end/models interface for a list of models to determine which models will work and which will first update the DSH itself to a new version.
Deepseek-v5 is the most basic model, does not state the client's version requirements, and the current version's built-in directory does not know it, so it is hidden and calls the Times MODEL UNSUPPORTED.
Delegates are open to all users. The declared minimum client version is 0.01, i.e. any version of the client can be used directly and will appear in the model list.
Deepseek-v5-next represents a subsequent updated model that requires an upgrade of DSH for use. The minimum version required was 99.0.0-alpha.1, while the currently public version of DSH remained at 0.2.x.
It is clear that DSH will never be updated to version 99.0.0, so Deepseek-v5-nextt is only a hand-written version of the test “never reach” to stabilize the branch that triggers the “no version”.
In other words, DeepSeek has written three possible scenarios for V5Z94Z in advance of the test: V5, which can be used without updating DSH, and V5, which is not known in the current version.
Some netizens believe that the models in the tests are fake data, which is just placeholders.
Although placeholders are indeed common in the software industry, they are used to pave the way for subsequent updates.
But the placeholder is essentially for engineers to turn off in advance. When V5 is on line, no matter which version of the DSH is installed by the user, there's no problem with the software.
Such a guarantee would be meaningful only if “V5 did come”. No one will design a mechanism for grade publish for a name not on the road map.
Besides, there are so many options, any string, and DeepSeek has chosen V5, which means either V5 is already in training, or V5 is already in the pipeline for the future.
And looking at the mechanism itself, you'll find that DeepSeek is trying to solve a “publish efficiency” problem.
Previously, the new model had to go online and the client had to issue a simultaneous version, and the user had to update the software to the new version before seeing it. Now, if the service side changes the return content of /models, the new model can be made directly accessible to the user, i.e., in the same way as deepseek-v5-open.
Changes to the “DeepSeek API key adapter now updates the model list in the same way as the account login” also appear in the warehouse's updated records, indicating that the mechanism is being extended to more entry points.
This means that DeepSeek V5Z94Z, except for DSH, other software with access to the DeepSeek model can be accessed in seconds to DeepSeek V5, and users need not be asked to update their client.
Deepseek-v5-nextt corresponded more to “new architecture, new capabilities”. If V5 requires new client capabilities, such as new reasoning parameters or new modes of interaction, this path can be useful.
The fact that both paths are available at the same time indicates that DeepSeek's specific approach to V5 is well thought out.
But let me add that these three V5 models are not uploaded today. The actual upload time was 24 September 2026, and the uploader was the morning of graduation from Qinghua University.
From the GitHub warehouse page, the morning is one of the core contributors to DSH. In the early stages of DSH, his submission was second in line, after Choi Tim Wing, and involved in a great deal of research and development, such as model adaptation, Web, and connection restoration.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.