Three model identifiers containing the label V5 have appeared in the test code of DeepSeek’s official Harness repository on GitHub, according to a TechNode report published on October 29. The strings are deepseek-v5, deepseek-v5-open and deepseek-v5-next. DeepSeek has not announced a V5 model, and the entries are test scenarios that contain no specifications, benchmark results or launch timing, the report says.

The find is a small technical detail with a large rumor attached to it. A model called V5 has been speculated about online since at least June, and independent trackers say it has not been confirmed. What the repository shows is that the name exists in code DeepSeek itself maintains, which TechNode presents as an early hint at naming in the company’s tooling, not as evidence of a release.

What the repository shows

TechNode, writing under its TechNode Feed byline, says the three identifiers sit in test cases for the software’s DeepSeek account integration. The tests check how the client handles model compatibility, including which models are listed as available and when the client needs to be updated. One test permits deepseek-v5-open to be listed next to DeepSeek Flash. The other entries produce compatibility-related outcomes that the article does not describe.

TechNode links to the file in question, a test specification in the deepseek-ai organization’s deepseek-harness repository, but the file itself could not be opened for this report, so the description rests on TechNode’s account. The article does not explain what the suffixes open and next mean, and it names no one at DeepSeek. This suggests caution: test fixtures are often written to exercise software logic and can include names that do not correspond to a shipped product, though the report does not say whether that applies here.

What DeepSeek Harness is

DeepSeek Harness, abbreviated dsh, is an open-source harness for AI agents, meaning the software around a language model that supplies tools, runs the agent loop and manages memory. Its GitHub page carries the tagline “Everything is a Plugin,” is published under the MIT license and describes itself as a developer preview in which compatibility-breaking changes should be expected. The repository page showed roughly 268,000 stars and 32,500 forks when it was opened for this report.

A guide posted on DEV Community on October 7 by Sergey Sereda says DeepSeek open-sourced the tool in August 2026 and counted about 245,000 stars in early October. The guide says the default provider is deepseek-official with deepseek-v4-flash as the default model, and that the tool also supports providers such as Anthropic, OpenAI, Kimi and GLM.

Where DeepSeek’s model line stands

DeepSeek’s own API changelog gives the official sequence. On April 24 the company released a V4 preview with two open-weight models. DeepSeek-V4-Pro has 1.6 trillion total parameters with 49 billion active, and DeepSeek-V4-Flash has 284 billion total with 13 billion active, both with a one-million-token context window. DeepSeek said V4-Pro was “rivaling the world’s top closed-source models” in math, STEM and coding.

A V4-Flash update on July 31 moved that model’s API to public beta, and V4-Pro reached general availability on August 13. DeepSeek released an experimental vision variant, deepseek-v4-flash-vision-exp, on August 21. On September 10 it released V4.1 Flash under the API name deepseek-flash, which the changelog calls the smallest model in a new architecture family and which retired the earlier V4 Flash models. The latest entry on the page, dated October 17, upgraded the model to V4.1-Flash-1017 and opened the company’s in-house search engine to API users. The same entry recommends Harness version 0.2.1-rc.1 or later.

The name deepseek-flash in the changelog matches the term DeepSeek Flash in TechNode’s description of the test. This suggests the tests concern the current Flash model, though TechNode does not make the link. The changelog entries opened list no V5 model.

The V5 rumor trail

Sources that track the rumor disagree on its origin. DataLearner, a model-tracking site, says V5 is a community nickname and not an official name, and that the speculation traces to a single YouTube channel, World of AI, in late August. It adds that as of August 31 DeepSeek had announced no V5 model, roadmap or release date, and that the next model could ship under a V4.x name. A June 26 blog post by Thomas Wiegold, an AI developer in Sydney, says there was no V5 as of June and already cites a video thumbnail promising a V5 launch on July 24, which suggests the speculation is older than DataLearner’s account.

The site Shattered.io, in a page by Daniel Roth updated on October 27, says that as of October 24 it found no changelog entry, no model identifier on the API and no technical report for a V5. It lists two triggers for confusion. One is DeepSeek’s notice that the older deepseek-chat and deepseek-reasoner aliases would be retired after July 24, which some readers took as a launch signal. The other is a September 2 forecast attributed to Bank of America Merrill Lynch that placed a possible V5 launch in the second half of 2026, which the site describes as a forecast and not a DeepSeek statement. That forecast could not be traced to a primary document.

The rumored September window passed without a V5 release, according to Shattered.io, which also says a V4.1 Pro model has not launched. All of these sites are independent third parties, not DeepSeek, and their claims about leaks and early testers are not independently verified.

Other signals: chips and money

Two other threads help explain why a new DeepSeek model draws attention. TechNode reported on July 8, citing a Reuters report from July 7 that relies on unnamed sources, that DeepSeek had started an in-house AI chip project aimed at inference, the work of serving trained models to users, rather than training. The goals, according to the report, are to lower inference costs with custom processors and to reduce reliance on overseas suppliers such as NVIDIA. Development reportedly began about a year earlier and hiring had ramped up in recent months, through targeted outreach to experienced chip engineers rather than public job postings. DeepSeek had not commented, and TechNode noted that the path from chip design to mass production typically takes well over a year.

On funding, a summary published October 22 by LetsDataScience says Bloomberg reported that DeepSeek had raised about 100 billion yuan, roughly $15 billion, in its latest round, with Reuters putting the figure at $14.98 billion and both naming CATL and Tencent among the largest backers. The same summary says the first outside round closed in June at about 50 billion yuan. It says Reuters cautioned that the figures could still change, and that DeepSeek has not confirmed them. Bloomberg and Reuters were not opened directly for this report.

What is not known

The repository entries leave basic questions open. Nothing in TechNode’s account says whether V5 is a new generation, a renamed point release or a stand-in name, whether the open and next variants are separate products, or when anything might ship. DeepSeek has made no statement, and the sources reviewed contain no comment from anyone at the company.

The release history offers only a partial guide. DeepSeek’s changelog shows five model updates between July 31 and October 17, with gaps of eight to 37 days between them. It does not say whether tooling changes precede model releases, and nothing in the sources reviewed establishes a timeline for V5.

What to watch

The most reliable signals are the ones the independent trackers name: a new entry in DeepSeek’s API changelog, a live model string on the API, and a technical report or model repository under the deepseek-ai organization. This suggests that further changes in the Harness repository, such as configuration or pricing for the new identifiers, would be firmer evidence than the current entries, which TechNode says contain no such details.

Until then, the accurate summary is narrow. According to TechNode, a repository DeepSeek controls contains three V5-labeled strings in compatibility tests. The company has said nothing about them, and its own changelog still shows the V4 family, with V4.1-Flash-1017 as the latest model.

Sources

Image: BalticServers data center.jpg by BalticServers.com, CC BY-SA 3.0, via Wikimedia Commons. Illustrative photo; it does not show the events described.