Image: @ZeroZ_JQ on X

Three model names that point to a DeepSeek V5 have allegedly been spotted in DeepSeek's own code. X user ZeroZ_JQ posted on October 28 that deepseek-v5, deepseek-v5-open and deepseek-v5-next appear in the latest code for DeepSeek Harness, and wrote that a release did not seem far off.

🐳鲸鱼兄弟们！DeepSeek 要放大招来啦！

V5、V5-open、V5-next都端上桌了。

deepseek Harness 的最新代码中发现已经有相关的代码了。

看来离发布也不远了。
github.com/deepseek-ai/deepseek-harness pic.x.com/9GfrWyjz1W

— 关木 (@ZeroZ_JQ) October 28, 2026

X user ItsmeAjayKV shared the find later that day, saying DeepSeek V5 was spotted in the Harness code and that it might arrive before the end of the year.

DeepSeek v5 is spotted in deepseek-harness code !

We might just get it before end of this year!🤩 pic.x.com/OxF0TJU3PP x.com/ZeroZ_JQ/status/2115285824300527713

— AJ (@ItsmeAjayKV) October 28, 2026

What the code actually shows

The names are real and sit in a public repository under DeepSeek's GitHub account. They appear in a test file for the part of DeepSeek Harness that signs users in with a DeepSeek account. DeepSeek Harness is the company's open source agent tool, described in its repository as a system where everything is a plugin.

The test checks how the app handles models listed by DeepSeek's server. In it, deepseek-v5-open shows up in the model list, deepseek-v5 is hidden and refused as unsupported, and deepseek-v5-next asks for an app upgrade to version 99.0.0-alpha.1. Public releases of the tool are still on version 0.2, so that version number looks like a value chosen to trigger an error in testing.

Other outlets have picked up the find. TechNode reported that the entries are part of test scenarios and contain no model specifications, benchmark results or release schedule. Chinese tech site TMTPost noted that some users see the names as placeholders, but argued that preparing different launch paths suggests V5 is on the roadmap. TMTPost also reported that the lines were added on September 24, which would mean they sat in the code for about a month before drawing attention.

TMTPost said the setup would let DeepSeek switch on a new model from its server without a client update, while still blocking models an older app cannot handle. It also pointed to recent repository changes that bring the API key route in line with the account route for reading model lists.

Why caution is needed

Test files often use made-up names to check how software reacts to new data, so these lines do not prove that DeepSeek V5 is finished or coming soon. They also say nothing about how V5 would perform or what it would cost.

DeepSeek has not announced V5 or commented on the code at the time of this writing. Its most recent release covered here was DeepSeek V4.1 Flash 1017, which arrived with open weights earlier in October. Until DeepSeek says more, the V5 names should be treated as an early hint rather than a confirmed product.