Deepseek v5 is here

DeepSeek V5 seems to have been built a while ago, but the official release hasn't happened yet.

This afternoon, Teacher Meng spotted a screen full of deepseek-v5 in the DeepSeek Harness code and sent me a screenshot.

1. It's a model name, not an API version

Teacher Meng searched the repository, and every occurrence of deepseek-v5 is used as a model name, not an API version number.

The evidence is in this line, from packages/llm/llm-deepseek/tests/loader-composition.spec.ts:169. This file simulates the response of https://api.deepseek.com/models:

{
  id: 'deepseek-v5',
  api_capabilities: [...],
  client_requirements: { ... }
}

id in an OpenAI-compatible interface holds the model name. When a client requests /models, the list will include deepseek-v5 alongside existing models, and it can be called directly with model="deepseek-v5". The adjacent api_capabilities and client_requirements are fields that describe it, not its name.

2. It was in the code on September 24th

Teacher Meng checked the git history. The name first appeared on 2026-09-24 13:06, commit d8c9868, authored by Yichen Jiang.

The commit message is feat(llm): gate DeepSeek account models by declared client version and bundled support—deciding which models are visible in an account based on the client's declared version.

This commit added resolve-catalog.ts, which parses the model catalog; tests are in resolve-catalog.spec.ts, using deepseek-v5 directly as an example. That afternoon this file was revised twice more, within five minutes.

So v5 plays the role of a model that exists in an account but is invisible if the client version is too old. The test verifies a single rule: when the server delivers models, it includes client_requirements.min_version, and if the client is too old, it doesn't appear in the list.

To test this rule, some model has to serve as an example; making one up on the fly would work. The test author used v5.

But in the first screenshot, v5 appears alongside the FLASH and PRO tier constants: { ...FLASH, id: 'deepseek-v5' }, { ...PRO, id: 'deepseek-v5' }. These test cases were adapted from the existing two models; a random name wouldn't be written across both tiers.

FLASH and PRO have been live for a while, and old clients can still use them—no need to block by version. So the only model this rule could block is a new model clients don't recognize yet: an unreleased model, v5.

Since that day, this name has been in the repository for over a month, and the model still hasn't arrived.

3. When will v5 actually arrive?

DeepSeek is backed by High-Flyer, a quantitative fund company, run by boss Liang. When to release a model is not just a technical question for High-Flyer. Every time DeepSeek ships a version, NVIDIA, compute rental, and that chain of A-share AI concept stocks all get a jolt.

So when is the best time to release?

Maybe Liang isn't refusing to release; he's just waiting for a window that can paint a beautiful K-line.