Three V5 Names in DeepSeek's Code, and Not One Model You Can Call

DeepSeek test code reportedly contains three V5 model identifiers, offering a possible clue to future development without establishing that any of the models is ready to use. ^[4]

The unconfirmed report names deepseek-v5, deepseek-v5-open and deepseek-v5-next, alongside deepseek-v4-pro, in tests for DeepSeek Harness. BigGo Finance attributes the discovery to IT之家 and describes the entries as framework-level placeholders rather than interfaces users can call. ^[4]

A separate, firmer development concerns DeepSeek's released V4.1 Flash model. Cognition's FrontierCode 2.0 results table gives it 19.3%, ahead of Kimi K3 at 17.7%. The benchmark tests whether AI-written changes are suitable for merging into code repositories. It supplies evidence about coding performance, not about the reported V5 names. ^[1]

The two developments deserve different weight. A benchmark score measures a model under a particular test setup. A name in a software test records a name that developers have made provision for. Treating both as evidence of an approaching V5 launch would turn a code clue into a product announcement.

Three names, no working interface

The V5 account rests on secondary reporting about the test code. BigGo reports that v5-open is associated with a client version requirement of 0.0.1, while v5-next is associated with 99.0.0-alpha.1. It interprets those values as suggesting different compatibility paths. ^[4]

A client is software that connects to a service. Version checks can determine whether that software meets a requirement before a feature is used. In tests, developers can use model names and version numbers to check how such rules behave without connecting to a working model.

That makes the reported entries worth watching, but weak evidence for a release schedule. The labels distinguish three cases in the framework. They do not explain what DeepSeek intends each case to become. In particular, the word open is not a licence, and next is not a launch date.

As of October 28, DeepSeek had announced no V5 model, published no model card describing its purpose, limits and licence, and provided no API identifier, the name developers use to call a model from their own software, according to BigGo. It had also made no public statement explaining the placeholders. The underlying test file has not been independently verified for this article, and the reported entries establish neither a trained V5 system nor its availability, capabilities or release timetable. ^[4]

The useful next evidence would be inspectable code showing the entries and their history, or documentation from DeepSeek explaining the names. Until then, the identifiers remain a reported development clue.

What the 19.3% score measures

V4.1-Flash-1017 has a much more concrete public record. DeepSeek's model card identifies it as a minor update to V4.1-Flash and states that both the repository and model weights carry the MIT licence. The weights are the downloadable files that contain the model's learned parameters. ^[3]

Pandaily's reported benchmark claim concerns that release. Cognition's own table lists the model under the shorter name DeepSeek V4.1 Flash, scoring 19.3% against Kimi K3's 17.7%, a lead of 1.6 percentage points. Cognition says each model is shown at its best reasoning effort, its setting for how much computing work a model can spend thinking before answering. ^[4][1]

That supports the stated comparison on FrontierCode 2.0. It does not establish that DeepSeek wins across programming languages, development tools or everyday software work.

Cognition designed the benchmark around whether an agent's changes could be accepted into a real code repository. An agent here is a system that uses a model to carry out coding work. The test therefore reaches beyond whether the model can produce a plausible answer to a detailed programming prompt. ^[1]

The scoring also makes the percentage easy to misread. FrontierCode 2.0 uses a weighted rubric, with solutions receiving zero if they fail blocking criteria. The 19.3% figure should not be read as a simple claim that the model completed roughly one in five ordinary coding jobs. It is a score produced by those rules. ^[1]

Cognition says the second version uses larger tasks and more blocking criteria than its predecessor. The aim is practical mergeability: whether proposed changes meet the requirements for becoming part of the codebase. ^[1]

An independent method is not a replicated result

Artificial Analysis describes its own implementation of FrontierCode 2.0 as an independent evaluation using 132 private tasks. Its runs use the Stirrup agent harness, the software framework that manages the agent's work, and attempt each task three times before averaging the scores. ^[2]

Artificial Analysis also gives zero to attempts flagged for looking up solutions. ^[2]

But Artificial Analysis's description does not establish an independent reproduction of the specific 19.3% figure in Cognition's table. That result remains Cognition's published measurement. The exact configuration, harness and run settings behind the DeepSeek entry would be needed for a close replication. ^[1][2]

There are therefore two distinct next steps. For the coding result, an independent run of the same model under comparable conditions would test whether the lead holds.

For V5, the next meaningful development is a verifiable code record or a statement from DeepSeek. Another model name, by itself, would still leave users without a model to run.