The account @ItsmeAjayKV says DeepSeek V5 code has shown up in the deepseek-harness codebase and that the release could come before the end of 2026. The post attaches one image and gives no further detail. Chinese-language and financial news reports describe the same find with more specifics.

Image shared by @ItsmeAjayKV on X
Image shared by @ItsmeAjayKV on X

Image credit: @ItsmeAjayKV on X

What the code contains

Reports name three identifiers: deepseek-v5, deepseek-v5-open and deepseek-v5-next. A 163.com report places them in a test file, provider.spec.ts, for the account login path of DeepSeek Harness. It says the client fetches a model list from the server's /models endpoint after login, and that the list decides which models are usable and which need a DSH update first.

An SMZDM article gives each identifier a minimum client version:

Identifier

Minimum client version (per SMZDM)

deepseek-v5

None declared

deepseek-v5-open

0.0.1

deepseek-v5-next

99.0.0-alpha.1

SMZDM says the missing version on deepseek-v5 means the current client hides it and calls return MODEL_UNSUPPORTED. It describes the 99.0.0-alpha.1 requirement as deliberately out of reach. The public build is still in the 0.2.x line, with the latest public release at 0.2.2-alpha.2.

KuCoin, carrying an AIBase report, says V5 is currently only a placeholder name and that DeepSeek could enable it with a server-side change to the /models configuration, without users updating their clients.

Timing and what remains rumour

SMZDM, citing a GitHub commit, says the three entries were added on September 24, 2026 by core contributor Yichen Jiang. That is about five weeks before they were noticed publicly on October 27–28. KuCoin says netizens spotted the code on the evening of October 27, after DeepSeek changed its model 'fingerprint' the same night.

KuCoin says the code does not confirm an immediate release and that a November launch looks more likely than October. SMZDM calls the November timing, and the idea that V5 Pro and V5 Flash would launch together, extrapolations from past update cadence rather than an official schedule. A Sina report says analysts expect V5 Pro and V5 Flash possibly in November. SMZDM also says its parameter-count and Huawei Ascend training claims are rumours it does not verify.

As of SMZDM's publication, DeepSeek had made no announcement, model card or API identifier for V5.

Our read: the reports say the placeholders sit in the repository, and KuCoin says DeepSeek could enable V5 with a server-side change to /models. The useful signal to watch is the /models list, not the placeholder names.

What to watch

Watch for V5 to appear in DeepSeek's live /models list or the app model picker, and for any official model card. Our earlier report on the leak covers the first reporting, and our Harness download guide explains what the current preview build supports.

Sources