For almost two years, a slice of the AI world held its breath for a model that had a name but no release date. The name was R2. The expectation was enormous. The product never came.
What DeepSeek actually put on the table in 2026 was quieter, stranger, and in some ways more interesting than the sequel everyone imagined.
The release that actually happened
On 24 April 2026, DeepSeek released a preview of its V4 large model (大模型), available as two variants — V4-Pro and V4-Flash — and, as with its earlier releases, under an open-weight licence that lets developers download, run and modify the code (CNBC, 24 April 2026).
The headline numbers, as the company and multiple reports describe them:
- V4-Pro: about 1.6 trillion total parameters, with roughly 49 billion activated per token, using a Mixture-of-Experts design.
- V4-Flash: about 284 billion total parameters, with roughly 13 billion activated.
- Both support a 1-million-token context window — enough to drop an entire codebase or a year of emails into a single prompt.
- Both ship under an MIT-style licence for self-hosting.
None of that is the rumored R2. It is the V-series, promoted.
Why "no R2" is the real story
R2 was cast as the successor to R1, the January 2025 reasoning model that rattled global markets. For months, specs for R2 circulated — parameter counts, benchmark scores, even a launch month — none of them confirmed by DeepSeek. Reporting from Reuters said founder Liang Wenfeng held the release because he was not satisfied with its performance, and no timeline was ever set.
As of mid-2026, R2 still does not appear in DeepSeek's official API model list. The legacy deepseek-chat and deepseek-reasoner endpoints, which had been quietly routing to V4-Flash, were scheduled to be retired after 24 July 2026.
The lesson is not that DeepSeek missed a launch. It is that the lab chose to absorb reasoning into the main model rather than ship a separate "thinking" product. V4 decides, per task, how hard to think: a simple question gets a fast answer; a hard proof triggers deeper deliberation. The R lineage now lives inside V4.
Trained where it had to be
One detail separates V4 from its predecessors. According to reporting by Reuters and The Information, sections of the model stack were rewritten for compatibility with Huawei Ascend silicon, and the model was trained on that hardware rather than Nvidia's. The export-control squeeze that helped stall R2 became the constraint that shaped V4's architecture.
For a lab whose previous headline was "frontier performance on weaker chips," training on domestic accelerators is less a political statement than a supply-chain fact. It also makes V4 a proof point for the rest of China's AI stack: a trillion-parameter-class model that does not depend on Washington's permission.
The pricing move nobody copied
DeepSeek kept its signature low price, and added a twist. V4 introduced time-of-day pricing — peak hours in Beijing (09:00–12:00 and 14:00–18:00) cost roughly double the off-peak rate. Off-peak API rates sit far below those of US frontier models. Treat compute like electricity: cheap when the grid is idle, expensive at rush hour.
We are deliberately not quoting a single cents-per-million-token figure here, because published numbers differ across sources and shift with the peak/off-peak schedule. The durable point is the posture: a frontier-adjacent model priced as a utility, not a luxury.
What it means for the open tier
DeepSeek's role in the ecosystem is to keep the downloadable frontier within touching distance of the closed one. Every time it ships a capable open-weight model at aggressive cost, it compresses the premium that closed labs can charge. V4, with its scale and context length, does exactly that — which is why the rest of the field watches the lab so closely, and why the absence of R2 barely registered in the market.
The competitive frame also changed. When R1 launched, DeepSeek was a lone disruptor. By April 2026, Alibaba, ByteDance and others had shipped their own strong open models, and DeepSeek's own positioning now pits Chinese open-weight labs against one another — a rivalry that did not exist when R1 dropped.
Honest limitations
- Parameter counts and the 1M context come from DeepSeek's release and were echoed across multiple secondary reports; we did not independently verify the architecture or benchmark the models.
- "Trained on Huawei Ascend" rests on Reuters / The Information reporting, not a primary DeepSeek training document we accessed.
- Pricing figures vary by source and by the peak/off-peak schedule; we describe the structure qualitatively rather than cite a single rate.
- R2's status is a mid-2026 snapshot. If DeepSeek later ships an R2, this piece documents the moment before it.
- Benchmark claims that circulate in aggregator write-ups are vendor or third-party figures, not our measurements.
What readers can do now
- Build on what shipped, not on rumors. If your plans rested on "R2," re-baseline against V4-Pro and V4-Flash, which are live and downloadable.
- Test the 1M context on a real document problem — legal review, research synthesis, long agent runs — where context length, not raw speed, is the bottleneck.
- If sovereignty or cost matters, pilot self-hosting under the MIT-style licence and compare token economics against your current closed model on your own workloads.