As of July 2026, DeepSeek R2 still has not launched, and DeepSeek has never confirmed a release date. The company that shook the AI world with R1 in early 2025 spent 2026 shipping something else entirely, the DeepSeek V4 series, while its next dedicated reasoning model stayed locked behind closed doors. Reports point to founder Liang Wenfeng holding R2 back because he was not satisfied with its performance.
This page tracks exactly where DeepSeek R2 stands right now, what has actually shipped versus what is still rumor, and the specs, pricing, and timeline the leaks point to. You will also see why the launch keeps slipping, the chip drama behind the delays, and the real question few are asking, whether R2 will even arrive as a separate model at all. Every confirmed fact is sourced, and every rumor is clearly flagged as unverified.
The Key Takeaways
- Not released. As of July 2026, DeepSeek R2 has no confirmed launch date and does not appear in DeepSeek’s official API.
- V4 shipped instead. DeepSeek previewed V4-Pro (1.6T params) y V4-Flash (284B) on April 24, 2026 under the MIT License, with the official version due mid-July 2026.
- Rumored specs: a roughly 1.2-trillion-parameter hybrid Mixture-of-Experts model, double R1’s size, trained on 5.2 PB of data. None of this is confirmed.
- Chip trouble delayed it. Training reportedly stalled on Huawei Ascend hardware, forcing a pivot back to Nvidia for training.
- DeepSeek is scaling up. A first-ever $7.5B funding round in June 2026 valued the company near $58.9B.
Is DeepSeek R2 Out Yet?
As of July 2026, DeepSeek R2 has not been released and has no confirmed launch date. DeepSeek previewed its V4 series, V4-Pro y V4-Flash, in April 2026 instead. Reports from Reuters say founder Liang Wenfeng held R2 back over performance concerns, so the timeline stays open.
The rumor mill has been wrong before. DeepSeek publicly denied a widely circulated March 17 release claim, and a string of “coming this month” predictions across 2025 and 2026 never landed. R2 still does not show up in DeepSeek’s official API model list, which currently serves deepseek-v4-pro y deepseek-v4-flash, not an R2 endpoint.
If you want a signal to watch, prediction markets like Polymarket and Manifold run live odds on an R2 launch, and they have repeatedly slipped their own resolution dates. The honest answer is that no credible confirmed date exists, and anyone claiming otherwise is reading tea leaves.
What Is DeepSeek R2?
DeepSeek R2 is the anticipated successor to DeepSeek R1, the reasoning model that shook the AI world in early 2025 by matching far more expensive Western models at a fraction of the cost. R2 is expected to be a dedicated reasoning model, one built to “think” through problems step by step rather than answer instantly. What actually arrived in 2026 was the V4 series, and MIT Technology Review broke down why V4 matters for the wider AI race.
The catch is that DeepSeek has confirmed almost nothing. Everything circulating about R2’s architecture, size, and capabilities comes from supply-chain leaks and analyst speculation, not official announcements. Treat the “specs” below as directional rumor, not fact.
DeepSeek R2 vs R1: What’s Rumored to Change
Compared with R1, DeepSeek R2 is expected to roughly double the parameter count to around 1.2 trillion in a hybrid Mixture-of-Experts design, add multimodal vision, and improve reasoning efficiency. R1 remains the shipping reasoning model today, and every R2 figure comes from leaks rather than DeepSeek, so treat the comparison as provisional.
| Feature | DeepSeek R1 (shipping) | DeepSeek R2 (rumored, unconfirmed) |
|---|---|---|
| Release | January 2025 | Not released as of July 2026 |
| Architecture | Mixture-of-Experts | Hybrid MoE with dense layers |
| Parameters | ~671B total | ~1.2T total (double R1) |
| Context window | 128K | Expected larger, up to ~1M |
| Modality | Text | Text plus vision |
| Training hardware | Nvidia | Huawei Ascend, then Nvidia (reported) |
| Focus | Reasoning | Stronger reasoning, efficiency |
| Status | Available | Unconfirmed |
Rumored Features and Specs
Here is what the leaks claim, with the caveat that DeepSeek has verified none of it. Read each point as “reportedly,” not “confirmed.”
- ~1.2 trillion parameters in a hybrid Mixture-of-Experts layout, roughly double R1 (unconfirmed).
- 5.2 petabytes of high-end training data spanning finance, law, and patents (unconfirmed).
- Multimodal vision capabilities, a step up from R1’s text-only focus (unconfirmed).
- Major efficiency gains, with some leaks touting dramatically lower inference costs (unconfirmed and, frankly, hype-adjacent).
- Advanced gating to route tokens to specialist experts more precisely than R1 (unconfirmed).
DeepSeek’s whole reputation rests on delivering frontier-level reasoning cheaply, so an efficiency-focused R2 fits the pattern. Just remember that “40x more efficient” style claims have circulated for over a year without a shipping model to back them up.
Why the Launch Keeps Slipping: The Chip Saga
The most concrete reason for R2’s delays is hardware. According to reporting from Tom’s Hardware, DeepSeek was urged by Chinese authorities to train R2 on domestic Huawei Ascend chips, and the runs kept failing. The Ascend 910C reportedly suffered stability problems, weaker inter-GPU bandwidth, and immature software.
The reported fix was a compromise, train R2 back on Nvidia hardware while using Huawei Ascend chips for inference. That kind of mid-project hardware pivot is exactly the sort of thing that blows up a release schedule, and it lines up with Liang Wenfeng’s reported reluctance to ship until the model is genuinely ready.
Is R2 Even a Separate Model, or Is It “V4-Thinking”?
Here is the question most trackers skip. DeepSeek’s V4 series already delivers strong reasoning, and prediction markets now literally frame the bet as “R2 / V4-Thinking.” That hints the market itself is unsure whether R2 will arrive as a standalone model or fold into a reasoning-tuned variant of V4.
It would not be unusual. The rest of the industry has blurred the line between “chat” and “reasoning” models, baking step-by-step thinking directly into flagship releases. If DeepSeek follows suit, the R2 name could quietly disappear, replaced by a thinking mode on top of V4 rather than a separate launch. Nothing is confirmed either way, but it is the scenario worth watching.
One near-term signal to watch: DeepSeek has said the official V4, graduating from its April preview, lands in mid-July 2026 with a new peak and off-peak pricing structure. The company has said nothing about a separate R2 shipping alongside it, which keeps the “R2 folds into V4” scenario very much alive.
DeepSeek’s Latest Moves in 2026
Whatever R2 turns out to be, DeepSeek is clearly building for scale. The moves it made through mid-2026 all point to a company gearing up for a much bigger model, not winding down.
A first-ever $7.5B funding round
In June 2026 the company closed its first external funding round, raising roughly $7.5 billion (about 51 billion yuan) at a valuation near $58.9 billion, breaking Liang Wenfeng’s long-standing “no fundraising, no IPO” stance. As the South China Morning Post reported, DeepSeek is signalling a clear push to train bigger models for less.
Data centers and a hiring spree
That capital is going somewhere. DeepSeek launched a hiring spree across seven business lines and accelerated data-center construction in Ulanqab, Inner Mongolia. Both are the kind of infrastructure investment you make ahead of a major model launch, not after one.
The DSpark inference breakthrough
On the research side, a June 27 paper co-authored by Liang introduced DSpark, an inference framework the team claims speeds up generation by 60% to 85% without changing the underlying model. Faster, cheaper inference is precisely the foundation a next-gen reasoning model like R2 would need.
DeepSeek R2 Pricing: What to Expect
There is no official R2 pricing, because there is no R2. What we can anchor to is DeepSeek’s track record of aggressive, category-low prices. The current DeepSeek V4 lineup sits far below Western frontier models on cost per token, and R1 built its reputation on undercutting OpenAI by an order of magnitude.
Expect R2, whenever it lands, to follow the same playbook, frontier-class reasoning at a price that pressures everyone else. For the current numbers, see our DeepSeek pricing breakdown, and check our roundup of the best open source AI models for where DeepSeek stands against its rivals today.
How to Use DeepSeek Today on Mac and iPhone
You do not have to wait for R2 to put DeepSeek to work. DeepSeek V4 is already live and, as an open-weight model, it runs in plenty of third-party apps beyond DeepSeek’s own site. If you would rather not route everything through a China-hosted service, that choice matters, because with the right app you decide which model sees each prompt.
That is where Fello AI fits in. It runs DeepSeek alongside GPT, Claude, Gemini, and Grok in one native Mac and iPhone app, so you can switch models mid-chat and compare answers without juggling five subscriptions. When R2 or a V4-Thinking mode finally ships, you get it in the same place, no vendor lock-in and no waiting on a single provider. To go deeper on the model that actually shipped, read our full DeepSeek V4 breakdown, and the story of founder Liang Wenfeng.
The Bottom Line
DeepSeek R2 remains a rumor, not a product. As of July 2026 there is no confirmed release date, no official specs, and a real chance the “R2” name gets absorbed into a reasoning-tuned V4. What is certain is that DeepSeek has the funding, the data centers, and the research pipeline to ship something big when it is ready.
Bookmark this page, we update it the moment R2 (or its successor) actually launches. In the meantime, the smartest move is to use DeepSeek V4 today and stay model-agnostic, so whatever DeepSeek ships next, you are ready to test it on day one.
FAQ
Is DeepSeek R2 out yet?
No. As of July 2026, DeepSeek R2 has not been released and does not appear in DeepSeek’s official API. The company shipped its V4 series in April 2026 instead.
When will DeepSeek R2 be released?
There is no confirmed date. DeepSeek has publicly denied earlier rumored dates, and reports say founder Liang Wenfeng is holding the model back until its performance meets his bar.
How is DeepSeek R2 different from R1?
Rumors point to roughly double the parameters (around 1.2 trillion), a hybrid Mixture-of-Experts design, added vision, and better efficiency. None of this is confirmed by DeepSeek.
Why does DeepSeek R2 keep getting delayed?
The main reported reason is hardware. Training reportedly stalled on Huawei Ascend chips, forcing a pivot back to Nvidia for training while using Ascend for inference.
How much will DeepSeek R2 cost?
There is no official pricing yet. Based on DeepSeek’s history of undercutting rivals, expect aggressive, category-low pricing similar to the current DeepSeek V4 lineup.
