DeepSeek R2 - rumor tracker

DeepSeek R2: release date, delays, and why it's taking so long

DeepSeek R2 is the brand's most anticipated and most rumored model: the direct successor to R1, the reasoning model that made DeepSeek famous in early 2026. Except R2 still doesn't officially exist. The company has denied several dates, and its CEO reportedly deemed the performance not good enough to ship. This page lays out where things stand, without inventing a date, and points you to what's actually available today for reasoning.

Your questions about DeepSeek R2

What's confirmed, what isn't

Factual answers to the most frequent searches about R2.

When is DeepSeek R2 coming?

No confirmed official date. Several release windows have circulated since early 2025, and DeepSeek has denied them each time. To date there is no official communication announcing a date for R2. Any date you see elsewhere is speculation. We update this page as soon as an official source moves.

Was DeepSeek R2 cancelled or just delayed?

Delayed, not cancelled, based on available information. According to Reuters, DeepSeek's CEO was reportedly not satisfied with R2's performance and preferred to hold the release rather than ship a disappointing model. Hardware difficulties for training were also mentioned. Nothing points to an abandonment: DeepSeek simply won't ship R2 until it hits its target.

What's the difference between R2 and R1?

R1 (early 2026) is the reasoning model that put DeepSeek on the map: visible chains of thought, ultra-low-cost training, performance close to the best at a fraction of the price. R2 is expected to push reasoning further, especially on code and multilingual tasks. But since R2 hasn't shipped, any numbered R2-vs-R1 comparison is speculative. Be wary of tables claiming otherwise.

Has R2 already been replaced by something else?

Partly, yes. Since R1, DeepSeek has built reasoning directly into its general-purpose models: V3.1 already merged the two approaches, and V4 offers adjustable thinking modes (with or without reasoning). In other words, much of what people expected from R2 is already available in V4-Pro and V4-Flash, with no need to wait for a separate R2 model.

Are the "32B, 92.7% on AIME" R2 specs real?

Unverified. Very precise numbers circulate about R2 (model size, benchmark scores), but none come from an official DeepSeek source. They're leaks or estimates copied from site to site. Until DeepSeek publishes a model card, treat these specs as noise, not information.

bell

What should I use for reasoning while waiting for R2?

DeepSeek V4-Pro in thinking mode. It already delivers very high-level reasoning (90.1% on GPQA Diamond in the sciences), with an adjustable thinking budget and a one-million-token context. For most of the tasks that fueled the wait for R2, V4-Pro is the best answer available at DeepSeek today.

Sorting facts from rumors

DeepSeek R2: what we can state

We won't give you a fake date to chase clicks.

Confirmed

  • R2 has no official release date
  • DeepSeek has denied several release windows
  • Reasoning is already built into V3.1 and then V4
  • V4-Pro in thinking mode covers most needs

Unverified rumor

  • Any precise R2 release date
  • The "32B" specs or AIME scores attributed to R2
  • The numbered R2-vs-R1 or R2-vs-o-series comparisons
  • The idea that R2 has been cancelled for good

While waiting for R2

DeepSeek reasoning is already here

V4-Pro in thinking mode already delivers what many expect from R2.

DeepSeek V4-Pro

Reasoning mode, 90.1% GPQA Diamond

Learn more

DeepSeek V4-Flash

Free, fast, adjustable thinking

Learn more

DeepSeek V5 rumors

The next generation, unconfirmed

Read the take
info

New model on July 24?

No. What actually changed

Understand it

Model timeline

From R1 to V4, the full story

See the timeline

DeepSeek V4 vs Gemini

Reasoning against Google's flagship

Compare

Don't wait

R2 has no date? DeepSeek reasoning is already available

No need to wait for R2 for high-level reasoning. Turn on V4-Pro's thinking mode and test it on a math, logic, or coding problem. It's the best DeepSeek answer today to the need that fuels the wait for R2.