DeepSeek R2: Release Status, Rumors & DeepSeek V4 (2026)

Last updated: June 2026 · By DecodeTheFuture.org

What DeepSeek actually released in 2026: V4, V4-Pro and V4-Flash

While the press chased R2 rumors, DeepSeek’s real 2026 flagship arrived on April 24, 2026: the DeepSeek V4 family, launched live on chat, app and API with open weights. R2 — a next-generation reasoning model — was not part of that release.

So if you are choosing a DeepSeek model today, the real options are V4-Pro, V4-Flash and the existing R1 reasoning line — not R2. R2, if and when it ships, would be the next reasoning model built on this V4-era base. The rest of this article explains the rumored R2 profile and why it generated so much expectation.

What is DeepSeek R2?

DeepSeek R2 is the anticipated second generation of DeepSeek’s reasoning-first model line — but it has not been released, so everything described here is rumored, not confirmed. R1 (January 2025) was a 671-billion-parameter Mixture-of-Experts model. The most-circulated R2 leak describes something very different: a 32B dense transformer under MIT license, small enough to fit on a single RTX 4090 or A6000. If that profile turns out to be accurate, it would be a dramatic pivot — but no official model card or weights exist to confirm it.

The rumor mill has been loud and contradictory, which is part of why expectations ran ahead of reality. Throughout 2025, leaks suggested R2 would be a 1.2-trillion-parameter MoE model trained on Huawei Ascend chips, then on Nvidia hardware after Ascend stability problems forced a pivot; later leaks flipped to the small dense design above. Reuters reported that CEO Liang Wenfeng held the release because he was not satisfied with R2’s performance. The net result: as of June 2026, R2 remains unreleased and its real architecture is unknown.

Why does the size drop from 1.2T to 32B matter?

It matters because it inverts the assumption that drove the entire post-GPT-4 era: that frontier reasoning requires hundreds of billions of activated parameters. R2 puts most of its intelligence into post-training — specifically a refined version of the GRPO reinforcement-learning pipeline DeepSeek introduced with R1 — rather than into raw scale.

The practical consequences for developers are immediate:

A 92.7% AIME 2025 score is not a casual benchmark. AIME — the American Invitational Mathematics Examination — is the qualifier round for the USA Math Olympiad. A score of 92.7% means R2 correctly answers roughly 14 out of 15 problems where each problem demands multi-step symbolic reasoning. For comparison, the original R1 hovered around 74% on the same benchmark in independent evaluations, and GPT-5’s reported scores sit in a similar range without tool use.

How would a 32B R2 reason like a 671B model? (the rumored approach)

This describes the approach the leaks point to — it is not a confirmed R2 spec. The short version: distillation plus a longer reinforcement-learning post-training phase, building on three techniques DeepSeek has genuinely been refining since R1.

1. Reasoning distillation from a larger teacher

The likely recipe: use the full R1 (and reportedly a high-end DeepSeek reasoning variant from late 2025) as a teacher model. The teacher generates millions of long chain-of-thought traces for math, code, and logic problems; the 32B student is then fine-tuned on those traces. This is the same playbook that produced the original R1-Distill-Qwen-32B back in January 2025 — but with eighteen months of accumulated technique on top.

If you want the full mechanics of this process, our explainer on LoRA fine-tuning covers the parameter-efficient side, and our piece on deep learning fundamentals walks through why distillation transfers reasoning patterns more efficiently than retraining from scratch.

2. GRPO with self-verification

Group Relative Policy Optimization (GRPO) was DeepSeek’s original RL contribution: instead of training a separate value network, you sample a group of responses to the same prompt, score them against a verifier, and update the policy toward the higher-scoring members of the group. R2 would reportedly layer self-verification on top — training the model to check its own intermediate reasoning steps before committing to a final answer, the direction DeepSeek’s math-focused research has been taking.

3. Dense, not sparse

This is the structural choice that makes everything else viable. A dense 32B transformer activates every parameter on every token, which means no expert-routing overhead, no load-balancing tricks, and — critically — no need for the 8-GPU minimum that MoE inference imposes. The trade-off is that you cannot scale dense architectures to trillion-parameter sizes the way MoE can. For a model that’s deliberately small, that ceiling doesn’t matter. Our transformer architecture explainer goes deeper on why dense vs. MoE is the central design fork in modern LLM engineering.

How much does DeepSeek R2 actually cost to run?

Two paths, two cost structures.

These are projections, not posted R2 prices — R2 is not released. Via an API, if R2 follows DeepSeek’s usual reasoning-model pricing, it would land well below GPT-5 and Claude 4.6. DeepSeek has historically priced reasoning models around $0.45–$0.55 per million input tokens and $2.00–$2.20 per million output tokens; frontier Western reasoning APIs sit closer to $3 / $15 per million for the highest tiers. For an agentic workflow burning 20 million tokens/day, that envelope is the difference between roughly a $40/day and a $250/day bill — which is why R1 and the new V4 family are already attractive today.

Self-hosted, a small dense architecture would change the economics entirely. If the rumored 32B profile is accurate, R2 could run at INT4 on a single RTX 4090 (24 GB VRAM, ~$1,800 used) at 30–40 tokens/second — the floor below which API economics stop making sense for high-volume, latency-tolerant workloads. Today that floor is set by released open models like the 32B R1 distill and the V4-Flash tier; an R2 in this profile would push it further.

# NOTE: DeepSeek R2 is not released — no "deepseek-r2" model exists yet.
# This is the pattern you would use once it ships. Until then, swap in a
# released model, e.g. "deepseek-r1:32b" or "deepseek-v4", which exist today.

from ollama import chat

response = chat(
    model="deepseek-r1:32b",  # placeholder: a real, released model
    messages=[{
        "role": "user",
        "content": (
            "A particle moves along the curve y = x^3 - 6x^2 + 9x. "
            "Find the values of x where the tangent line is horizontal. "
            "Reason step by step, then give the final answer in \\boxed{}."
        ),
    }],
    options={"temperature": 0.6, "top_p": 0.95},
)

print(response.message.content)

Note the temperature=0.6 and the explicit “reason step by step” instruction — both inherited from DeepSeek’s recommended R1 settings. R2 still benefits from this scaffolding; without it, the model occasionally truncates its own chain of thought.

How does R2 compare to GPT-5, Claude 4.6, and Gemini 3.1 Pro?

The honest answer is: nobody knows, because R2 has not been released and no independent harness (Artificial Analysis, MathArena, LMArena) has tested it. If R2 ships near its rumored profile, this is the likely shape based on DeepSeek’s track record with R1 — read every row as a projection, not a measured result:

The picture that emerges is the same one R1 painted in January 2025: R2 is not the absolute best at anything, but it is the cheapest entry point to “good enough” reasoning by a margin that nothing else in the open ecosystem comes close to matching.

What does R2 mean for the open-source AI ecosystem?

If R2 ships near its rumored profile, three shifts would follow — and the first two are already visible from R1 and the V4 release.

First, the economic case for closed reasoning APIs gets harder. If you build a product that does math, code review, or structured analysis at scale, a cheap open reasoning model lets you move that workload off GPT-5 or Claude 4.6 and onto DeepSeek’s API or your own GPU. R1 and the V4 family already enable this today at a fraction of frontier-API cost; an R2 in the rumored profile would widen the gap further.

Second, the geopolitics tighten. R2 was reportedly trained on Nvidia hardware after the Huawei Ascend pivot failed in 2025 — a quiet acknowledgment that domestic Chinese AI silicon is still 18–24 months behind. U.S. export-control hawks will use this to argue for tighter restrictions; Chinese policymakers will point to results like the rumored 92.7% AIME score to argue that the restrictions don’t work. Both arguments are partially correct.

Third, the distillation playbook becomes the dominant strategy. A small dense R2 of the rumored kind would be a vindication of the idea that you train a giant teacher once, then distill it into many small specialized students — a pattern R1’s own distilled variants already demonstrated. Expect every major lab to publish 7B / 14B / 32B distilled versions of their flagships within the next quarter, with reasoning quality that would have been unthinkable a year ago.

What does this mean in practice for developers in 2026?

The deeper shift R2 reveals is that the dominant constraint on useful AI is no longer raw model capability — it’s the cost of running that capability against your actual data. We’ve crossed into an era where the relevant question is not “which model is smartest” but “which model is smart enough at a price I can sustain”. For most reasoning workloads in mid-2026, that question already has strong open-weight answers — R1 distills and the V4 family today, with a rumored small dense R2 potentially pushing the bar further. The implication for anyone building AI agents or RAG pipelines is that the model layer is rapidly becoming a commodity — and the durable engineering value is moving up the stack into context engineering, retrieval quality, and evaluation infrastructure.

That’s not bad news. It means smaller teams with sharper engineering can compete on equal terms with labs that have ten thousand times the capital. DeepSeek’s open line — R1 and V4 today, R2 if it ships as rumored — keeps making that competition more affordable.

How to access DeepSeek reasoning today (and what to do when R2 ships)

You cannot use DeepSeek R2 yet — it has not been released, there are no R2 weights on Hugging Face and no R2 model id in any API. Anyone offering “DeepSeek R2” access today is serving a mislabeled or fine-tuned stand-in. Here is what actually works now, and what to watch for when R2 arrives.

Use DeepSeek’s released models now

For reasoning today, the real options are DeepSeek R1 (deepseek-reasoner) and the new DeepSeek V4 family (deepseek-v4-pro, deepseek-v4-flash), all OpenAI-compatible. You can call them via DeepSeek’s own API, or via multi-model providers (OpenRouter, Together, Fireworks, the NVIDIA NIM catalog). For how hosted-API economics work — free tiers, rate limits, 429 errors, NIM vs OpenRouter — see our NVIDIA NIM API pricing & limits guide.

from openai import OpenAI

# use a model that actually exists today, e.g. V4 or R1
client = OpenAI(base_url="https://api.deepseek.com/v1", api_key="YOUR_KEY")

resp = client.chat.completions.create(
    model="deepseek-v4-pro",  # or "deepseek-reasoner" (R1)
    messages=[{"role": "user", "content": "Prove there are infinitely many primes."}],
)
print(resp.choices[0].message.content)

What to check when R2 is actually released

• An official DeepSeek announcement and a model card (not a third-party blog or leak).
• An R2 model id appearing in the official API model list.
• Open weights published on DeepSeek’s Hugging Face org with a stated license.
• Independent benchmarks (Artificial Analysis, MathArena, Vals) — not just vendor-reported scores.