DeepSeek V5 names appear in test code as V4.1-Flash tops open models on coding

What turned up in the code

As pandaily.com reported, developers browsing DeepSeek's public Harness project on GitHub noticed model names that were not there before: deepseek-v5, deepseek-v5-open and deepseek-v5-next. Harness is DeepSeek's own software for working with its models. Pandaily checked the repository itself and found the names in a test file, sitting at lines 154 to 156 beside an existing deepseek-v4-pro entry.

The repository history ties them to a commit dated September 24. That change moved the DeepSeek account's model catalog over to the official model listing. In other words, the names arrived as part of housekeeping, not as part of a launch.

What the test actually checks

The test looks at how the client program filters which models a user sees, based on the version of the client. The v5-open entry requires a minimum client version of 0.0.1, so practically any client qualifies. The v5-next entry requires 99.0.0-alpha.1, a version number no real client would have. The test expects only the flash model and v5-open to appear in the resulting list.

Pandaily's reading is that this shows Harness can turn a new model on from the server side. DeepSeek could reveal a model to every client, or only to clients that have upgraded, without shipping a fresh release of the software. That is a statement about plumbing, not about what DeepSeek is building.

Why this is not a V5 announcement

The source is blunt on this point. The entries are placeholders in test code; nobody can call them as models. Some developers think the names are nothing more than stand-ins used to exercise the filtering logic, and the code says nothing about whether a V5 exists or when it might ship.

As of October 28, DeepSeek had put out no announcement, no model card (the document describing what a model is and how it was built) and no API identifier for a V5 series. The current lineup is V4-Pro plus V4.1-Flash. The latter came out in September with a new underlying design but kept the V4 name, a reminder that DeepSeek's version numbers do not always signal what changed inside.

A real result: the coding benchmark

The same week brought firmer news. Artificial Analysis, which tracks and ranks AI models, released Intelligence Index v5, its biggest overhaul since v4 landed in January. Now 60% of the weighting rests on private test sets or hidden answers, which makes it harder for a model to look good just by having seen the questions. Coding makes up 25% of the index, and 15 points of that come from Cognition's FrontierCode 2.0.

FrontierCode 2.0 uses 132 real tasks from open-source projects. A model's proposed code change is judged by whether the project's maintainers would accept it, which is closer to real work than a puzzle with a known answer.

DeepSeek V4.1-Flash-1017, run at its maximum-effort setting, scored 28%. That is the best result among models whose weights have been released, meaning ones you can download and run yourself. Kimi K3.1 Pro posted 25.5%, but its full weights are not out yet. Gemini 4 Argon reached 25.7% and GPT-6 Luna 25.2%, so the open-weight model finished ahead of those listed rivals on this one test.

The less flattering overall picture

On the full index, the story is different. V4.1-Flash scores 32, level with GLM-5.3, while the leader, Claude Opus 5.5, scores 58. Artificial Analysis also says the distance between the best released open-weight model and the top of its table grew from 3 points under the previous version of the index (v4.3) to 12 under v5.

So a model can lead its category on one demanding coding test and still sit well behind the top overall. A single benchmark measures one slice of work, and the reweighting toward hidden test sets seems to have made the open-versus-closed difference more visible.