DeepSeek V5 Release Date: What We Know, and Why V4.1 Pro Comes First

DeepSeek V5 is not out, and DeepSeek has not announced it. The newest DeepSeek release is DeepSeek-V4.1-Flash-1017, an update shipped on 17 October 2026, and the next model DeepSeek has named in public is V4.1-Pro, not V5. The most specific report on V5 timing comes from the Chinese financial outlet Caiwen. On 3 November 2026 it said training is still being planned and pointed to a release sometime in 2027.

That has not stopped a steady stream of V5 claims.

There is a viral “leak”, there are model names in DeepSeek’s own code, and there is a new report that V5 will be trained on Huawei chips instead of Nvidia. This guide checks each claim against DeepSeek’s own pages and the original reporting, explains what the code actually says, and tells you which DeepSeek model to use while you wait.

The Key Takeaways

• Not released: DeepSeek’s news page and API changelog have no V5 entry. The latest model is V4.1-Flash-1017 (17 Oct 2026).
• Next in line: DeepSeek itself said V4-Pro traffic would run on V4.1-Flash “until V4.1-Pro launches”. V4.1-Pro has not shipped yet either.
• Timing: Caiwen’s industry sources point V5 at 2027, with training still in the selection phase. An older Bank of America forecast of H2 2026 now looks out of date.
• Huawei: Two Chinese outlets reported on 3 Nov that V5 will be pretrained on Huawei Ascend chips. Neither DeepSeek nor Huawei has confirmed it.
• The “leak”: “deepseek-v5” in DeepSeek’s Harness code is a test placeholder. Two tests give it a context window of 9 tokens.

Is DeepSeek V5 Out?

No. DeepSeek V5 has not been released, previewed or announced. DeepSeek’s API changelog, the dated record of every model it ships, ends on 17 October 2026 with the DeepSeek-V4.1-Flash-1017 update and a new search API. The news page on deepseek.com lists V4.1-Flash, released on 10 September 2026, as its latest model. DeepSeek’s Hugging Face account, where it publishes open weights, has no V5 repository.

Two models are live on the DeepSeek API today. The table shows what each name serves.

When Is DeepSeek V5 Coming Out?

DeepSeek has given no date for V5. The best available answer is 2027 at the earliest, based on the one report that describes where the project stands. Every date in circulation comes from a forecast, a report citing unnamed sources, or a guess, so each one is listed here with who said it.

The Most Specific Report Points to 2027

The most specific report puts DeepSeek V5 in 2027. Caiwen (财闻) reporter Zhang Xiaodi cited people in the chip supply chain on 3 November 2026. They said DeepSeek V5 will use the Huawei Ascend 950 as its core training compute, and that the project is aimed at next year. The Caiwen report says V5 training is still in the selection phase. The parameters, features and release date are internal secrets. Its sources expect a release in the first or second half of 2027, depending on how training goes.

A model still in the planning stage does not ship in weeks. Pretraining alone takes months, and post-training and testing come after that. Even a smooth run that starts late in 2026 puts a public V5 well into 2027.

The Bank of America Forecast for 2026

The earliest date anyone printed came from Bank of America. 36Kr summarised the bank’s China AI monitoring report on 2 September 2026. It said “DeepSeek V5 and the upgraded version of ByteDance’s Seed series are expected to debut in the second half of 2026”. That is an analyst expectation, not a report from inside DeepSeek.

With less than two months of 2026 left and V4.1-Pro still unreleased, the Caiwen timeline is the one to plan around.

Why “November” Keeps Coming Up

The November 2026 date traces back to Kuai Keji (快科技), a Chinese tech site. On 28 October it reported V5 model names in DeepSeek’s code and guessed that a V5 Pro would not arrive in October but would come in November. The code it found is real. The November date was the writer’s own estimate, and the code itself does not support it.

DeepSeek’s Next Model Is V4.1-Pro, Not V5

The only future model DeepSeek has named in public is V4.1-Pro. When it launched V4.1-Flash on 10 September 2026, DeepSeek called it “the smallest model in our new architecture family”, designed for “scaling to larger models”. It then said it was phasing out V4-Pro and would route that traffic to V4.1-Flash until a bigger sibling arrived. DeepSeek’s official account posted the plan word for word:

⚡ V4.1-Flash is now live on the DeepSeek API with native multimodal support.

Set your model to deepseek-flash.

🔹 V4-Flash & V4-Flash-Vision-Exp are retired. For compatibility, deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash.
🔹 Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We’re phasing out V4-Pro.
🔹 Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches.

🤝 Official partners @WorkBuddy_AI (including Codebuddy) & @opencode now fully support V4.1-Flash. Try it today!

4/6

— DeepSeek (@deepseek_ai) September 10, 2026

DeepSeek later changed course. Its changelog entry for 10 September now adds that “in response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026”. So V4-Pro-0813 is still on the API, and V4.1-Pro has not appeared. The V4.1-Flash model is a 552-billion-parameter mixture of experts with 8 billion active parameters for input and 16 billion for output, according to DeepSeek. A V4.1-Pro would be the larger model in that same family.

The naming matters for anyone waiting on “the next DeepSeek”. DeepSeek’s own stated order is V4.1-Pro first. A V5 would be a new generation after it, which fits the 2027 reporting.

The DeepSeek V5 “Leak” in DeepSeek’s Own Code

The “deepseek-v5” names in DeepSeek’s code are test placeholders, not evidence of a finished model. DeepSeek Harness is the company’s open-source coding agent, published at github.com/deepseek-ai/deepseek-harness. We read the five test files where the names appear, at the version merged on 4 November 2026 for the Harness 0.2.2 release candidate.

What the Test Files Actually Do

The tests check how the Harness app treats a model it has never seen before. DeepSeek’s server sends the app a list of available models. Each test feeds in a fake list and checks whether the app offers the model, refuses it, or tells the user to upgrade. The fake unknown model in those lists is called “deepseek-v5”. Listed without a minimum client version, plain deepseek-v5 is refused as unsupported.

The variants have jobs, not specs. “deepseek-v5-open” is the unknown model the app is allowed to serve. “deepseek-v5-next” demands Harness version 99.0.0-alpha.1, a version that does not exist, so the test can confirm the app asks for an upgrade.

In two of the test files, the code sets deepseek-v5’s context window to 9 tokens. No real model has a 9-token context window.

What the Code Does Tell You

The tests show that DeepSeek built the Harness to accept new models without an app update. When a new model appears in the server’s list, the app can offer it straight away. That makes a future launch smoother. It says nothing about when the launch happens, what the model is called, or how big it is. The placeholder had to be named something, and “v5” is the obvious next number.

DeepSeek V5 and Huawei Ascend Chips

The newest V5 claim is about hardware. STAR Market Daily (republished by 21Jingji) and Caiwen reported on 3 November 2026 that DeepSeek plans to pretrain V5 on Huawei Ascend chips instead of Nvidia GPUs. Neither DeepSeek nor Huawei has confirmed it. Pretraining is the long, expensive stage where a model learns from raw data. Huawei chips have so far mostly run finished DeepSeek models, a job called inference.

How the Reporting Changed in Two Months

The reporting on DeepSeek’s use of Huawei chips for training moved in three steps in two months. Each report adds detail, and none of them comes from DeepSeek.

The Information’s September scoop came straight from its own account:

Exclusive: DeepSeek CEO Liang Wenfeng told investors that training AI models on Huawei chips is one of the company’s biggest bets and “has to work.”

DeepSeek expects a new batch of Huawei training chips as early as the fourth quarter.

Full story: https://thein.fo/4rn0vWd

— The Information (@theinformation) September 21, 2026

What DeepSeek Has Shipped for Huawei Chips

DeepSeek’s Huawei work is public code, even if the V5 plan is not confirmed. On 30 September 2026 DeepSeek published Ascend versions of its core libraries on GitHub. Its FlashMLA attention kernels reach up to 410 TFlops on the Ascend 950 during prefill, which DeepSeek puts at 95% of the hardware’s peak. DeepGEMM-Ascend and DeepEP-Ascend port its matrix and communication libraries, and the TileKernels project now runs the same Python code on Nvidia GPUs and Huawei NPUs. The TileKernels page thanks Huawei for its “technical support and engineering expertise”.

Training is the hard part that code alone does not prove. CLS noted that a Shenzhen research team had already completed full post-training of the 1.6-trillion-parameter V4-Pro on Ascend 910C chips, and stressed that this did not include pretraining. CLS called a V5 pretrained on Ascend the first real test of whether a top Chinese model trains on Chinese chips. Until now, it said, that question had drawn only doubts.

DeepSeek V5 Rumors, Graded

Most V5 claims online trace back to a few reports, and several of them were stretched along the way. The table below traces each one to where it started.

The Viral DeepSeek V5 “Leak” Post

The best-known example is a viral post from 26 September 2026 that bundled several reports into one “leak”:

🚨 DeepSeek V5 Leak: Beats Astra

>DeepSeek is reportedly preparing an imminent V5 launch
>Founder Liang Wenfeng calls it the company's biggest bet yet
>Rumored at 2 trillion parameters (not 3T)
>Reportedly the first DeepSeek model to train fully on Huawei Ascend chips instead of Nvidia
>Needs roughly 4x more Ascend accelerators than the Nvidia equivalent to hit the same training scale

DeepSeek is reportedly keeping the open-weight strategy

— Priya (@glitch_catAI) September 26, 2026

Every DeepSeek V5 Claim, Checked

None of the eight DeepSeek V5 claims in circulation is confirmed by DeepSeek, and only the Huawei training plan comes from reporting about V5 itself. Each one is checked here against its original source.

How DeepSeek Releases Models

DeepSeek has shipped new generations more than a year apart and point releases every few months. Its changelog gives exact dates, and those dates are a better guide than any leak.

V3 to V4 took 16 months. A V5 in 2027 would put a similar gap after V4, and would still leave room for V4.1-Pro first. DeepSeek has also warned readers about rumors before. Its V4 announcement included a line worth keeping in mind: “please rely only on our official accounts for DeepSeek news. Statements from other channels do not reflect our views.”

The reasoning model is a separate story. DeepSeek’s reasoning line stopped at R1. The long-rumored follow-up has its own page: our guide to DeepSeek R2 and its release rumors covers why it never shipped as a separate model.

What to Use While You Wait for DeepSeek V5

DeepSeek-V4.1-Flash-1017 is the model to use today, and it is free in the DeepSeek app and web chat. DeepSeek says V4.1-Flash beats the older V4-Pro on performance, cost and speed, so most people have no reason to wait. The free DeepSeek chat has no paid tier to upgrade to.

On the API, DeepSeek’s pricing page lists the current model at $0.15 per million input tokens and $0.60 per million output tokens off-peak. Peak rates double to $0.30 and $1.20. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. V4-Pro costs $0.66 and $1.98 off-peak. Both models take a 1-million-token context.

Our DeepSeek pricing guide has the full rate card, including cache-hit prices.

V4.1-Flash also ships open weights under the MIT license on Hugging Face, like every recent DeepSeek model. For background on the generation it replaced, read our DeepSeek V4 explainer. To see how DeepSeek stacks up against Qwen, Kimi and GLM, start with our ranking of the best Chinese AI models.

On a Mac, DeepSeek has no official chat app. Its chat app is for iPhone and Android, and its only Mac download is the DeepSeek Harness agent app for Apple silicon Macs. Our DeepSeek client for macOS explains the options. Fello AI puts DeepSeek in the same chat window as ChatGPT, Claude and Gemini, so you can check a DeepSeek answer against another model without opening a new tab.

The Bottom Line

Do not expect DeepSeek V5 in 2026. The only report that describes the project says training is still being planned and points to 2027, and DeepSeek’s own stated order puts V4.1-Pro first. The “leak” in the code is a test placeholder, and the Huawei training plan is a credible but unconfirmed report. Use V4.1-Flash now. Watch three places for the real announcement: DeepSeek’s API changelog, its Hugging Face account and the @deepseek_ai account on X. When V5 is real, it will show up there first.

FAQ