DeepSeek V5
DeepSeek hasn't said a word about V5 — but the runway is now clear. V4 Pro and V4 Flash came out of open-weight preview after the API's legacy deepseek-chat / deepseek-reasoner aliases retired July 24: our sweep read deepseek-v4-flash as official on July 31 and deepseek-v4-pro on August 13. V4 GA has landed. On Sep 10 DeepSeek shipped V4.1-Flash, the first model of a new architecture family, and named a larger V4.1-Pro as its next release; no V5 id has appeared in the catalogs we sweep. We watch the API either way: the second anything new answers, ArtificialWatch fires.
Leaks & news, newest first.
The “V5 leak” going around: no V5 exists yet
Posts are circulating that a DeepSeek V5 “leaked”, beat GPT-6 Astra, and was trained entirely on Huawei chips, with open weights. None of them cites a source, and as of Sep 27 there is no V5 model card, no weights and no API id in any catalog we sweep. What is real is the Sep 21 investor-meeting report below: a plan to train future models on Huawei hardware, not a finished V5. This page fires the moment a V5 id answers. via our own catalog sweep · The Information
Liang: training on Huawei chips is “one of our biggest bets”
At a closed-door investor meeting, founder Liang Wenfeng said DeepSeek plans to train its next models on Huawei and other domestic chips at scale, calling it one of the company’s biggest bets and one that must succeed. Per the report, Huawei’s Ascend 950 training chips could arrive as early as Q4 2026 or Q1 2027, and Liang estimated a frontier-scale run at roughly 50,000 Nvidia GB300s or 200,000 Ascend chips, about four to one. The meeting also discussed a planned 8-trillion-parameter model. No V5 name or date was given. via The Information · Android Headlines · Pandaily
V4.1-Flash opens a new architecture family — and names V4.1-Pro next
DeepSeek released V4.1-Flash, “the smallest model in our new architecture family”: a 552B MoE with 8B active parameters on input and 16B on output, native image understanding, and open weights under MIT. It first said V4-Pro calls would route to V4.1-Flash from Sep 14 “until V4.1-Pro launches,” then reversed that: V4 Pro keeps serving after Sep 14 at unchanged billing, and the old deepseek-v4-flash id now routes to V4.1-Flash. V5 was not mentioned. via DeepSeek · SiliconANGLE
V4-Pro reaches GA, with three effort levels
“The GA release of DeepSeek-V4-Pro has been rolled out on the APP, Web, and API,” with native Responses API support and low, high and max thinking effort. Peak/off-peak pricing followed on Aug 16, with off-peak at half the peak rate. via DeepSeek API docs
V4-Flash leaves preview
DeepSeek’s changelog: “The official release of the DeepSeek-V4-Flash API is now in public beta.” Same architecture and size as the April preview, with refreshed post-training. via DeepSeek API docs
The API housecleaning: legacy aliases retire July 24
DeepSeek's docs set July 24, 15:59 UTC as the retirement of the legacy deepseek-chat and deepseek-reasoner aliases — the kind of housecleaning labs do right before a stable release. V4 GA first; V5 is the follow-through. via DeepSeek API docs
V4 Pro and V4 Flash drop in preview — open weights, MIT
Both variants shipped to Hugging Face under MIT: V4 Pro at 1.6T parameters (MoE, ~49B active per token) and V4 Flash at 284B (~13B active). Preview status, no stable date announced. via DeepSeek · Hugging Face
V5: officially nothing — and that's the honest answer
No release date, no roadmap entry, no teaser. Every “V5 date” you've seen is extrapolation from DeepSeek's release cadence. When it exists, this page hears it first. via Chat-Deep roadmap check
The official word.
"Introducing the smallest model in our new architecture family, with native visual understanding."
DeepSeek, “Introducing DeepSeek-V4.1-Flash” · Sep 10, 2026 — the next model it names is V4.1-Pro; V5 has no official word
Until DeepSeek V5 ships: DeepSeek V4.
Which model is current until DeepSeek V5 ships?
The current model in this line is DeepSeek V4 from DeepSeek, live since April 24, 2026.
$0.66 / $1.98 per 1M tokens on DeepSeek's API.
Hear it the second V5 answers.
We sweep the public catalogs (OpenRouter, models.dev) and DeepSeek's own announcements every 60 seconds — the moment a V5 id answers, this watch fires. Email + Chrome push free, SMS and a phone call on paid tiers.