DeepSeek R2 release date: not released and no date announced (status as of October 13, 2026). DeepSeek’s official API docs, news page, changelog and Hugging Face organisation list no R2 model, no R2 API ID and no R2 weights. The last R-branded release was DeepSeek-R1-0528 (May 28, 2025). Since V3.1 (August 2025), DeepSeek has shipped reasoning as a “thinking mode” inside its V-series models. Today that means DeepSeek-V4.1-Flash (September 10, 2026, API name deepseek-flash) and DeepSeek-V4-Pro (GA August 13, 2026, deepseek-v4-pro). DeepSeek’s next announced model is V4.1-Pro, with no date given. Reports that R2 was planned for May 2025 and then delayed came from the press, not DeepSeek. Leaked R2 specs remain unconfirmed.
How to read this page. We label every claim as Official (published by DeepSeek in its API docs, news posts or Hugging Face repos), Reported (news organisations citing unnamed sources) or Rumour (leaks and social posts with no verifiable source). DeepSeek itself warned in its V4 announcement: “please rely only on our official accounts for DeepSeek news.”
Is DeepSeek R2 released? Status as of October 13, 2026
No. DeepSeek R2 has not been released, and DeepSeek has never published an R2 release date. We checked every official channel on October 13, 2026:
| Official channel | What it shows (Oct 13, 2026) | R2? |
|---|---|---|
| API docs: Models & Pricing | Two models: deepseek-flash (DeepSeek-V4.1-Flash) and deepseek-v4-pro (DeepSeek-V4-Pro-0813) | No |
| API news page | Latest posts: V4.1-Flash (Sep 10), V4-Flash-Vision-Exp (Aug 21), V4-Pro GA (Aug 13), V4 Preview (Apr 24, 2026) | No |
| API changelog | V4.1-Flash release; V4-Pro service continued after Sep 14, 2026 | No |
Hugging Face deepseek-ai | Newest repos: DeepSeek-V4.1-Flash, V4-Flash-Vision-Exp, V4-Pro-0813, V4-Flash-0731 | No |
Anyone selling “DeepSeek R2 API access” or offering “R2 weights” today is serving a renamed or fine-tuned stand-in, not an official DeepSeek model.
DeepSeek R2 release date timeline: what was reported vs what actually shipped
The R2 story is mostly a gap between press expectations and what DeepSeek actually released. This timeline puts both side by side:
| Date | Event | Type |
|---|---|---|
| Jan 20, 2025 | DeepSeek-R1 released: 671B MoE reasoning model, MIT license, API name deepseek-reasoner | Official |
| Early 2025 | Press reports say R2 was originally expected around May 2025 | Reported |
| May 28, 2025 | DeepSeek-R1-0528 update, still under the R1 name | Official |
| Jun 26, 2025 | Reuters, citing The Information: R2 launch stalled because CEO Liang Wenfeng was not satisfied with its performance | Reported |
| Aug 14, 2025 | Financial Times (as relayed by TrendForce and others): R2 delayed after unsuccessful training runs on Huawei Ascend chips; Nvidia used for training | Reported |
| Aug 21, 2025 | DeepSeek-V3.1: “one model, two modes” (Think and Non-Think). deepseek-reasoner now points to V3.1’s thinking mode | Official |
| Sep 29 / Dec 1, 2025 | DeepSeek-V3.2-Exp, then V3.2 and V3.2-Speciale | Official |
| Apr 24, 2026 | DeepSeek-V4 Preview: V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), 1M context, open weights | Official |
| Jul 24, 2026 | Legacy API names deepseek-chat and deepseek-reasoner scheduled to be fully retired (15:59 UTC) | Official |
| Jul 31 / Aug 13, 2026 | Official V4-Flash (0731), then V4-Pro GA (0813) with low / high / max thinking effort | Official |
| Sep 10, 2026 | DeepSeek-V4.1-Flash: first model of a “new architecture family”; DeepSeek mentions a coming V4.1-Pro | Official |
| Oct 13, 2026 | Still no R2 announcement, API ID or weights | Official (absence) |
What DeepSeek released instead: V4, V4-Pro and V4.1-Flash
While the press chased R2, DeepSeek kept shipping V-series models with reasoning built in. Here is the official lineup and API naming as of October 13, 2026:
| API model name | Serves | Size (official) | Context / max output | Price per 1M tokens (peak; off-peak is half) |
|---|---|---|---|---|
deepseek-flash | DeepSeek-V4.1-Flash (Sep 10, 2026), with vision | 552B MoE; 8B active for input, 16B for output | 1M / 384K | $0.30 input (cache miss), $0.006 cache hit, $1.20 output |
deepseek-v4-pro | DeepSeek-V4-Pro-0813 (GA Aug 13, 2026), text only | 1.6T total / 49B active MoE | 1M / 384K | $1.32 input (cache miss), $0.044 cache hit, $3.96 output |
deepseek-v4-flash, deepseek-v4-flash-vision-exp | Legacy names, temporarily routed to V4.1-Flash | The underlying V4-Flash models are retired; billed at the Flash price | ||
deepseek-reasoner, deepseek-chat | Legacy names (R1 / V3 era) | Scheduled for full retirement on July 24, 2026; no longer in the model table | ||
Both current models support thinking (the default) and non-thinking modes, JSON output, tool calls, the OpenAI Chat Completions and Responses formats, and an Anthropic-compatible endpoint. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. The V4-Pro and V4.1-Flash weights on Hugging Face are MIT-licensed.
Two details matter for anyone waiting on R2. First, the September 10 post calls V4.1-Flash “the smallest model in our new architecture family”, designed for “scaling to larger models”. It also says V4-Pro requests would route to V4.1-Flash “until V4.1-Pro launches”. DeepSeek’s changelog then reversed the routing: “we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026.” So the next officially signposted DeepSeek model is V4.1-Pro, not R2, and it has no date either. Second, V4.1-Flash is already a strong reasoner. DeepSeek reports GPQA Diamond 90.9, a Codeforces rating of 3471 and HLE 36.8 (vendor-reported).
For self-hosting requirements, see our DeepSeek V4 GPU and inference guide. For how sparse expert models like these work, see Mixture of Experts explained.
Is DeepSeek V4 the same as R2? The “DeepSeek V4 R2 release 2026” confusion
No. DeepSeek V4 (April 24, 2026) and V4.1-Flash (September 10, 2026) are general V-series models with a thinking mode. DeepSeek has never called any of them R2. The confusion has a structural cause. In August 2025, V3.1 merged the separate reasoning line into a single hybrid model (“one model, two modes”), and the R1 API name deepseek-reasoner became just the thinking mode of the V-series model. Since then, every DeepSeek release has followed that pattern.
Our interpretation, not a DeepSeek statement: because reasoning now ships inside each V release, a stand-alone “R2” model may never appear under that name. DeepSeek has not confirmed this either way. Until it does, the only accurate answer to “when is the DeepSeek R2 release date?” is that there is none.
The DeepSeek R2 rumours, rated
| Claim you will see online | Where it comes from | Status (Oct 13, 2026) |
|---|---|---|
| R2 was planned for May 2025, then delayed | Press reports (Reuters / The Information, Financial Times) | Reported, never confirmed by DeepSeek |
| CEO Liang Wenfeng held R2 back over performance | Reuters citing The Information, June 2025 | Reported |
| R2 training failed on Huawei Ascend; Nvidia used instead | Financial Times, August 2025 | Reported |
| R2 is a 1.2T-parameter MoE trained on Ascend | 2025 social-media leaks | Rumour, unconfirmed |
| R2 is a 32B dense model scoring 92.7% on AIME 2025 | Leaks without a model card, paper or weights | Rumour, unconfirmed |
| “R2 launched in March 2026” or “R2 API is live” | Third-party blogs and resellers | False: no official release, ID or weights |
Why vendor numbers need a second source. Even for released models, treat AIME-style scores as upper bounds until independent harnesses such as MathArena or Artificial Analysis publish their own runs. For R2 there is nothing to test: no weights, no API and no official benchmark table.
What R2 would change if it ships
This section is analysis of the rumoured profile, not a specification. If DeepSeek ever releases a separate R2, the most-discussed version is a much smaller, distilled reasoning model. Three ideas behind that rumour are real techniques DeepSeek has published:
- Reasoning distillation. A large teacher generates long chain-of-thought traces, and a smaller student is fine-tuned on them. DeepSeek already did this in January 2025 with the R1-Distill models, such as R1-Distill-Qwen-32B. See our LoRA fine-tuning guide for the parameter-efficient side.
- GRPO reinforcement learning. Group Relative Policy Optimization, which DeepSeek introduced with DeepSeekMath and used to train R1, scores groups of sampled answers against a verifier instead of training a separate value network. Our RLHF explainer compares it with PPO and DPO.
- Cheap local inference. A dense model of about 32B parameters fits on a single 24 GB GPU at 4-bit quantization, so it would compete with V4.1-Flash on self-hosting cost rather than raw capability.
The practical consequence is the same with or without an R2 label: open-weight reasoning keeps getting cheaper. DeepSeek’s own V4.1-Flash pricing ($0.30 input and $1.20 output per million tokens at peak) is already roughly 7 to 17 times cheaper per token than closed models such as Claude Sonnet 5.5 ($2 / $10) or Claude Opus 5.5 ($4 / $20). For a cross-provider view, see our LLM inference cost comparison.
How to use DeepSeek reasoning today (instead of waiting for R2)
Use the thinking mode of DeepSeek’s current models. This is the official OpenAI-compatible pattern from DeepSeek’s API docs, with deepseek-flash (V4.1-Flash) or deepseek-v4-pro:
import os
from openai import OpenAI
# There is no "deepseek-r2" model. Use a current DeepSeek model in thinking mode.
client = OpenAI(api_key=os.environ["DEEPSEEK_API_KEY"],
base_url="https://api.deepseek.com")
resp = client.chat.completions.create(
model="deepseek-flash", # DeepSeek-V4.1-Flash; or "deepseek-v4-pro"
messages=[{"role": "user", "content": "Prove there are infinitely many primes."}],
reasoning_effort="high", # low / high / max
extra_body={"thinking": {"type": "enabled"}},
)
print(resp.choices[0].message.content)
If your code still calls deepseek-reasoner (the old R1 name), switch it now. DeepSeek scheduled that name for full retirement on July 24, 2026. You can also reach DeepSeek models through multi-model gateways; our Ollama Cloud vs OpenRouter comparison and free LLM API guide cover the trade-offs.
What to check when R2 (or V4.1-Pro) is actually released
- A post on DeepSeek’s official API news page and changelog, not a third-party blog or leak.
- A new model name in the official Models & Pricing table.
- Open weights in the
deepseek-aiHugging Face organisation, with a license and tech report. - Independent benchmarks (MathArena, Artificial Analysis), not only vendor-reported scores.
FAQ
What is the DeepSeek R2 release date?
There is none. As of October 13, 2026, DeepSeek has not released R2 and has not announced a release date. Its official API docs, news page, changelog and Hugging Face organisation list no R2 model.
Has DeepSeek R2 been released in 2026?
No. In 2026, DeepSeek released the V4 Preview (April 24), the official V4-Flash (July 31), V4-Pro GA (August 13), V4-Flash-Vision-Exp (August 21) and V4.1-Flash (September 10). None of them is R2.
Is DeepSeek V4 the same as DeepSeek R2?
No. V4 and V4.1 are V-series models with a built-in thinking mode. DeepSeek merged reasoning into its V-series with V3.1 in August 2025, and it has never called any V4 model R2.
What is the latest DeepSeek model in October 2026?
DeepSeek-V4.1-Flash, released September 10, 2026 and called via the API name deepseek-flash. DeepSeek-V4-Pro (version 0813) remains available as deepseek-v4-pro. DeepSeek has said a V4.1-Pro is coming but has given no date.
Is the 32B / 92.7% AIME DeepSeek R2 leak real?
It is unconfirmed. No model card, paper, weights or official benchmark table for R2 exists. Earlier leaks described a completely different 1.2-trillion-parameter design, so treat any R2 parameter count or score as rumour.
Can I use DeepSeek R2 through the API?
No. There is no R2 model ID. For DeepSeek reasoning today, call deepseek-flash or deepseek-v4-pro with thinking mode enabled through DeepSeek’s OpenAI- or Anthropic-compatible API.
What happened to deepseek-reasoner (DeepSeek R1)?
From April 24, 2026, deepseek-reasoner pointed to the thinking mode of V4-Flash. DeepSeek scheduled it for full retirement on July 24, 2026, and it no longer appears in the official model table. The R1 weights remain on Hugging Face under the MIT license.
Sources
Status note: checked on October 13, 2026. “Official” facts come from DeepSeek’s API docs, news posts, changelog and Hugging Face repositories. Press reports are labelled as reports; leaked R2 specifications are unconfirmed. Benchmark figures are vendor-reported.
- DeepSeek API Docs: Models & Pricing (current model names, versions, prices, peak hours) · api-docs.deepseek.com/quick_start/pricing
- DeepSeek API Docs: Change Log (V4.1-Flash, V4-Pro GA, V4-Flash 0731, V4 Preview, legacy names) · api-docs.deepseek.com/updates
- DeepSeek: “DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient” (Sep 10, 2026) · api-docs.deepseek.com/news/news260910
- DeepSeek: “DeepSeek-V4-Pro GA Release” (Aug 13, 2026) · api-docs.deepseek.com/news/news260813
- DeepSeek: “DeepSeek V4 Preview Release” (Apr 24, 2026) · api-docs.deepseek.com/news/news260424
- DeepSeek: “DeepSeek-V3.1 Release” (Aug 21, 2025; hybrid think / non-think) · api-docs.deepseek.com/news/news250821
- Hugging Face: DeepSeek-V4.1-Flash model card and tech report (MIT) · huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- Hugging Face: DeepSeek-R1 model card (MIT) · huggingface.co/deepseek-ai/DeepSeek-R1
- DeepSeek-AI: “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning” (arXiv 2501.12948) · arxiv.org/abs/2501.12948
- Reuters: “DeepSeek R2 launch stalled as CEO balks at progress, The Information reports” (Jun 26, 2025; reported, not confirmed by DeepSeek) · reuters.com
- TrendForce: DeepSeek R2 launch reportedly delayed amid Huawei Ascend chip hurdles (Aug 14, 2025; summarises FT and Reuters reports) · trendforce.com

