DeepSeek V5 Release Date: The Viral “V5 Leak” Is Just Test Code (Oct 2026)

Short answer: there is no DeepSeek V5 release date. DeepSeek has not announced V5, teased it, or listed it anywhere official. The “V5, V5-open and V5-next are coming” posts going around this week trace back to a unit test in the DeepSeek Harness repo. Those three names are fake test data, added on September 24, and they say nothing about a launch. DeepSeek’s latest official release is V4.1-Flash-1017 from October 17, and the API today serves only deepseek-flash and deepseek-v4-pro. If someone gives you a specific DeepSeek V5 release date, treat it as a rumor. Everything below was checked against DeepSeek’s own docs and the raw code on GitHub on October 28, 2026.

Where the rumor came from: a test file

On October 28, around 03:30 UTC, a post on X said DeepSeek was about to drop something big. It showed code from the deepseek-harness repo and linked to line 154 of provider.spec.ts. It passed 60,000 views within hours.

The file path gives it away: tests/provider.spec.ts. The test is called “hides and refuses listed models this build cannot serve.” It fakes a response from the /models endpoint, then checks that the client hides the models it shouldn’t show. deepseek-v5, deepseek-v5-open and deepseek-v5-next were made up for that check.

The clearest tell is the client version deepseek-v5-next asks for: 99.0.0-alpha.1. DeepSeek Harness is on 0.2.x right now. Version 99 is deliberately absurd, so the test hits the “please upgrade your client” branch.

I also went through the file’s commit history. The V5 names first show up in a commit from September 24, 2026, titled “serve the DeepSeek account catalog from the account listing.” That’s a feature change, not a model announcement. Teortaxes, who follows DeepSeek closely, said the same: these are assert tests from late September, and V5-next isn’t coming any time soon.

What DeepSeek has actually released

DeepSeek posts every new model on the news page of its API docs. Here is 2026 so far:

• April 24: V4 Preview, with V4-Pro and V4-Flash open-sourced.
• August 13: V4-Pro general release.
• August 21: V4-Flash-Vision-Exp, an experimental multimodal model.
• September 10: V4.1-Flash, the smallest model in a new architecture family.
• October 17: V4.1-Flash-1017, the current latest.

The API quick start page lists two model names: deepseek-flash (served by V4.1-Flash-1017) and deepseek-v4-pro. No V5.

If you want a real hint, it’s in the September 10 post: the new architecture is built for “scaling to larger models.” So a bigger model is planned. DeepSeek just hasn’t said what it’s called or when it ships, so the honest DeepSeek V5 release date right now is: unknown.

How to tell when V5 is really coming

These are real signals. Screenshots and “I heard” are not.

• An announcement on the API news page. DeepSeek always posts one, with a date and a model name.
• A new name from the /models endpoint. Call the list models API with your own key instead of trusting someone’s screenshot.
• A new row on the pricing page. The official pricing page updates when a model goes live.
• Weights on Hugging Face. DeepSeek’s open models go up under its official account.

Model names in test code or config files don’t count. Plenty of teams invent future names to test their version checks.

FAQ

Does “V5-open” mean V5 will be open source?

No. It’s a made-up name in a test that checks client versions. It has nothing to do with licensing. Most of DeepSeek’s main models have shipped open weights, but DeepSeek hasn’t said anything about V5.

Someone says the DeepSeek V5 release date is next month. Should I believe it?

Not without an official source. DeepSeek usually announces a model on the day it goes live and rarely gives dates in advance.

V4.1-Flash-1017 or V4-Pro, which should I use now?

DeepSeek’s October 17 post says 1017 scores higher than the previous Flash and writes shorter answers, and it’s billed at the Flash price. Start with Flash for everyday coding and try Pro on harder tasks. Check the pricing page for current rates.

What to use while you wait

Waiting for V5 is no reason to sit still. For coding, DeepSeek says V4.1-Flash-1017 was tuned for Harness. My setup steps and real costs for Claude Code are in DeepSeek API for Claude Code, and the free vs paid options in OpenCode are in OpenCode DeepSeek Flash price. For how DeepSeek stacks up against other providers, see the LLM API pricing comparison.

I’ll update this page as soon as there’s an official DeepSeek V5 release date.