DeepSeek-R1 opens reasoning

The event and the narrow claim

On 20 January 2025, the Chinese laboratory DeepSeek released DeepSeek-R1, a large reasoning model, together with its model weights and six smaller distilled variants, all licensed for free commercial use. DeepSeek's own release announcement states that the code and weights are published under the MIT License and that the model reaches "performance on par with OpenAI-o1." The accompanying technical report makes a narrower and more specific claim: that the reasoning behaviour of a language model can be incentivized through reinforcement learning applied directly to a base model, without first collecting human-written reasoning traces.

That narrow claim is distinct from the broad cultural event that came to be called "the DeepSeek moment." The release itself was an engineering and licensing decision; the market panic and the debate about training costs that followed a week later were a separate phenomenon. A careful history keeps the two apart: what the paper demonstrated under stated conditions, and what observers inferred about the economics of frontier AI. This report treats the documented method and evidence first, and the downstream reaction as its own, more contestable, chapter.

The constraint: capable reasoning, but closed

The immediate predecessor was OpenAI's o1, previewed in September 2024. OpenAI described o1 as a model trained with reinforcement learning to "think before it answers," producing a long internal chain of thought before the final response, and reported large gains on competition mathematics such as AIME. o1 established that spending more computation at inference time on an extended reasoning trace could raise accuracy on hard problems. But o1's chain of thought was deliberately hidden from users, its weights were not released, and its training recipe was not published.

The constraint DeepSeek-R1 addressed was therefore not "can a model reason," which o1 had already shown, but access and inspectability. Before R1, the strongest reasoning behaviour lived behind a closed API whose mechanism was described only in general terms. R1's contribution was to make a comparable capability downloadable and its training procedure legible. That reframing matters: the event is less a first-of-its-kind capability than the opening of a capability that had been demonstrated but sealed.

The predecessor idea beneath both was older. Chain-of-thought prompting, described by Wei and colleagues in 2022, showed that simply asking a sufficiently large model to produce intermediate steps improved complex reasoning. o1 and R1 differ from that by training the model to generate those steps under reinforcement learning rather than eliciting them through a prompt.

The lineage inside DeepSeek

R1 did not appear from nothing. It was post-trained on DeepSeek-V3-Base, and the efficiency of that base model is part of why the release landed as hard as it did. The DeepSeek-V3 technical report describes a Mixture-of-Experts model with 671 billion total parameters, of which only 37 billion are activated per token, pretrained on 14.8 trillion tokens. Two architectural ideas carried over from DeepSeek-V2: DeepSeekMoE, a sparse expert design, and Multi-head Latent Attention (MLA), which compresses the key-value cache into a latent vector to cut inference memory.

The reinforcement-learning method also had a documented origin inside the lab. DeepSeekMath, published in early 2024, introduced Group Relative Policy Optimization (GRPO), a variant of Proximal Policy Optimization that estimates the advantage from a group of sampled outputs instead of a separately trained value network, reducing memory use. GRPO is the optimizer that R1-Zero and R1 use. So the "DeepSeek moment" rests on a stack of prior, incremental work: an economical MoE base, a memory-frugal attention scheme, and an RL algorithm the same team had already validated on mathematics.

R1-Zero: reinforcement learning with no warm start

The most striking experiment in the report is DeepSeek-R1-Zero. Here the team applied GRPO directly to the base model with no supervised fine-tuning at all, using simple rule-based rewards: one signal for whether a final answer was correct (checkable for maths and code), and another for whether the output followed a required format placing its reasoning between designated tags. There was no human-labelled reasoning data and no learned reward model.

Under this pure-RL regime, the report describes reasoning behaviour emerging on its own: the model learned to generate longer chains of thought, to re-examine and verify its own steps, and to allocate more thinking to harder problems. The authors highlight an "aha moment" where the model spontaneously reconsiders an approach mid-solution. Quantitatively, they report R1-Zero's pass@1 on AIME 2024 rising from 15.6% to 71.0% over training, and reaching 86.7% with majority voting over 64 samples, figures self-reported in the paper.

R1-Zero also showed the limits of the pure approach. Its outputs suffered from poor readability and language mixing, jumping between languages within a single answer. The behaviour was capable but not well suited to being used directly, which motivated the fuller recipe.

R1: the multi-stage recipe

DeepSeek-R1 proper is the pipeline built to fix R1-Zero's rough edges while keeping its gains. The report describes four stages. First, a small amount of high-quality "cold-start" data, a few thousand long chain-of-thought examples, is used to fine-tune the base model so its reasoning is readable from the outset. Second, large-scale reinforcement learning is applied to reasoning-heavy tasks, with an added reward that discourages language mixing. Third, once the RL model has converged, it is used to generate new data through rejection sampling, which is combined with supervised data for writing, factual question answering, and other general tasks. Fourth, a final reinforcement-learning stage aligns the model across all scenarios, balancing helpfulness and harmlessness with reasoning.

The design point is that cold-start data and RL do different jobs. The supervised warm-start supplies human-preferred structure and legibility that pure RL did not produce on its own; the RL supplies the capability gains that supervised imitation alone would not reach. This is why the report presents R1-Zero and R1 as a contrast rather than as two names for one thing: the former isolates what RL can do unaided, the latter is the deployable system that layers structure around it.

The numbers, and who reported them

On the headline benchmarks, the report positions R1 alongside OpenAI's o1-1217. It gives R1 a pass@1 of 79.8% on AIME 2024 and 97.3% on MATH-500, a Codeforces rating around the 96th percentile, 90.8% on MMLU, and 71.5% on GPQA Diamond, all documented in the technical report. These are task-specific results on specific evaluation sets, not a measure of general intelligence, and several are close enough to o1's reported figures that ranking depends on the exact test and decoding settings.

Two cautions belong with these numbers. First, they are self-reported by DeepSeek; independent replication on held-out or contamination-controlled sets is the appropriate follow-up for any such claim, and competition benchmarks like AIME are vulnerable to training-data contamination. Second, "on par with o1" is a statement about a set of quantitative benchmarks at one moment, not about every capability a user might care about. The strength of the evidence is that the weights are public, so third parties can and did re-run these tests, which is a materially different situation from a closed model's self-report.

The paper's later peer-reviewed version in Nature, published in September 2025, added external scrutiny that the original preprint did not carry.

Distillation to small dense models

Alongside R1, DeepSeek released six dense models fine-tuned on roughly 800,000 reasoning samples generated by R1, built on existing open bases: the DeepSeek-R1-Distill variants at 1.5B, 7B, 14B, and 32B on Qwen and at 8B and 70B on Llama. The claim in the release is that the 32B and 70B distilled models are competitive with OpenAI's o1-mini on reasoning benchmarks, with the report giving the 32B variant 72.6% on AIME 2024 and 94.3% on MATH-500.

The more interesting finding is comparative. The report states that distilling R1's outputs into a small model produced better reasoning than running large-scale reinforcement learning directly on that same small model. In other words, for a 32B model, imitating a strong reasoner was more effective than teaching it to reason from scratch with RL, which is expensive at that scale. This has a practical consequence that helped drive adoption: a laboratory or company without frontier-scale compute could obtain a usable reasoning model by distillation, using downloadable weights, rather than by owning the RL pipeline.

The competing approaches of early 2025

R1 arrived in a crowded field. OpenAI's o1 line was the reference point, and OpenAI had already previewed its next o3 generation in December 2024. Alibaba's Qwen team had shipped an open experimental reasoning model, QwQ-32B-Preview, in November 2024, reporting 90.6% on MATH-500 and 50.0% on AIME. Google had a Gemini "thinking" variant, and Anthropic later shipped extended thinking in Claude. Most directly, Moonshot AI's Kimi k1.5 technical report appeared in the same week, describing its own recipe for scaling reinforcement learning with long context for reasoning.

The approaches diverged on method as much as on openness. Some efforts used process reward models or Monte Carlo tree search to supervise intermediate steps; DeepSeek explicitly reported that these more elaborate techniques were less successful in its hands than simple rule-based rewards at scale, an unusually candid negative result. The competitive lesson of early 2025 was not that one recipe won, but that reinforcement learning against verifiable rewards had become the shared paradigm for eliciting reasoning, pursued in parallel by several labs.

The market reaction, and the cost it did and did not measure

A week after the release, DeepSeek's assistant app reached the top of the US App Store's free chart, and on 27 January 2025 several large-cap technology stocks sold off sharply, with Nvidia recording one of the largest single-day market-value losses in US history. The trigger was a narrative that a Chinese lab had matched frontier reasoning at a small fraction of the assumed cost.

That narrative rested on a conflation the DeepSeek-V3 report does not support. The often-quoted figure, roughly 2.788 million H800 GPU-hours costing about 5.6 million US dollars at an assumed 2 dollars per GPU-hour, is the report's estimate for the final pre-training run of V3. The report states plainly that it excludes the costs of prior research, ablations, and data. It is not the cost of creating R1, which adds a reinforcement-learning post-training stage on top of V3-Base, and it is not the cost of the hardware, the failed experiments, or the salaries behind either model. The efficiency claim is real and significant at the level it was made, the marginal cost of one training run, but it was widely repeated as if it were the total cost of building a frontier reasoning system, which it is not.

Downstream influence and the Nature milestone

The most concrete downstream effect was replication. Because the weights and method were open, GRPO-style reinforcement learning against verifiable rewards spread quickly into open-source reasoning projects, and R1's distilled checkpoints became common baselines and fine-tuning starting points. The recipe's central claim, that a base model can be pushed toward reasoning with rule-based rewards and minimal labelled data, was cheap enough to test that many groups did so within weeks.

A second effect was institutional. In September 2025, DeepSeek-R1 was published in Nature after peer review, an unusual step for a frontier large language model, most of which are documented only in company preprints and system cards. Peer review does not certify a model's safety or its benchmark supremacy, but it subjected the training claims and evaluation methodology to external referees, which is a higher evidentiary bar than a self-published report. For a field where "technical reports" often function as marketing, that is a meaningful precedent.

How to place the DeepSeek moment in AI history

The defensible reading is layered. The narrow, well-documented claim is that reinforcement learning with simple verifiable rewards can incentivize reasoning in a large model without supervised reasoning traces, demonstrated by R1-Zero, and that a multi-stage recipe with cold-start data turns that into a deployable system, demonstrated by R1. The medium claim is that comparable reasoning became openly downloadable under a permissive license, which broadened who could build on it. The broad claim, that this single release reset the economics of AI, is the least secure and the most repeated, and it depends on a cost figure that describes one pre-training run rather than the whole enterprise.

Three questions help place it against its neighbours in the timeline. Which capability did it open rather than invent (inference-time reasoning, already shown by o1)? Which method did it reuse and make legible (GRPO, from its own DeepSeekMath)? And which caveat survives (self-reported benchmarks, contamination risk, and a training cost that was real but partial)? Held to those questions, the DeepSeek moment is best understood as an access and transparency event layered on top of a genuine efficiency result, not as the day the cost of intelligence collapsed.

References

• DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.
• DeepSeek (2025). DeepSeek-R1 Release.
• DeepSeek-AI (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning.
• DeepSeek-AI (2024). DeepSeek-V3 Technical Report.
• DeepSeek-AI (2024). DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model.
• Shao et al. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.
• OpenAI (2024). Learning to Reason with LLMs.
• Kimi Team (2025). Kimi k1.5: Scaling Reinforcement Learning with LLMs.
• Qwen Team (2024). QwQ: Reflect Deeply on the Boundaries of the Unknown.
• Wei et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.
• DeepSeek-AI (2025). DeepSeek-R1 Model Weights.