DeepSeek-R1 reaches Nature after review by eight experts

On September 17, 2025, Nature published the DeepSeek-R1 technical work after its editorial and review process. The original source supports the documentary core of the event; review supports publication of the study, not every commercial generalization about the model.

Publication does not automatically turn all of DeepSeek’s claims into incontrovertible facts. It does, however, subject the methodology and results to scientific scrutiny that major US labs—accustomed to sharing their advances through technical reports and corporate presentations—have rarely accepted for their most capable models.

Reasoning emerged without human examples in R1-Zero

The paper’s central finding is that certain reasoning capabilities can be encouraged through reinforcement learning, a technique in which a system receives rewards when it reaches correct answers, without having to be taught step by step how to reason.

DeepSeek applied this method to R1-Zero, an experimental version built on DeepSeek-V3-Base. Rather than training it on reasoning written by people, the company used problems whose solutions could be checked automatically, particularly in mathematics and programming.

During the process, behaviors emerged such as reviewing an answer, trying another strategy, or spending more steps on difficult problems. This matters because those patterns were not introduced as human demonstrations: they emerged as a useful way to maximize the reward.

There is an important distinction. The DeepSeek-R1 released for general use was not trained exclusively through pure reinforcement learning. The company added a small initial phase using supervised examples, followed by reinforcement learning, the generation and filtering of new data, supervised fine-tuning, and a second reinforcement-learning phase. That additional process was intended to improve readability, reduce language mixing, and extend capabilities beyond problems with verifiable answers.

The paper therefore supports the idea that reasoning can emerge through pure reinforcement learning in R1-Zero—not that the entire final version of R1 does without supervised data.

What the $294,000 cost actually represents

The $294,000 figure refers to the reinforcement-learning training described as part of the development of its reasoning capabilities. It is not the total cost of creating DeepSeek-R1 from scratch, nor does it include all the research, discarded experiments, data preparation, or training of the base model. Document supporting the figure.

R1 builds on DeepSeek-V3, a mixture-of-experts architecture with 671 billion total parameters, of which around 37 billion are activated to process each piece of text. This technique distributes the model across numerous specialized blocks and uses only a portion of them for each operation, reducing the computation required compared with activating all parameters at once. Document supporting the figure.

DeepSeek had estimated the cost of V3’s final pretraining run at $5.576 million, calculated from 2.788 million hours of Nvidia H800 GPU time. That figure also did not represent the project’s full budget, as it excluded earlier development and other trials. Document supporting the figure.

The $294,000 figure is therefore significant but limited in scope: it shows that adding reasoning behavior through reinforcement learning can be far cheaper than pretraining the base model. It does not mean that R1 in its entirety cost that amount, nor does it provide a direct comparison with models such as OpenAI o1, whose labs have not published an equivalent breakdown. Document supporting the figure.

More open than usual, but not fully open

In January, DeepSeek released R1’s weights and several distilled versions—smaller models trained to imitate the larger system—under the MIT license. It also published a technical report describing its methodology and results, enabling researchers and companies to run the models on their own infrastructure.

Publication in Nature adds another layer: independent reviewers were able to challenge the experimental design, request clarifications, and assess whether the conclusions were supported by the evidence presented. This sets the work apart from model cards or safety reports produced and published solely by the companies themselves.

It does not, however, amount to an independent reproduction of the training. Nor does it make R1 a completely open project: having the weights does not necessarily provide the original data, all the training code, or the complete experiment history needed to reconstruct the model from scratch.

Pressure shifts to other labs

The paper’s significance extends beyond DeepSeek. OpenAI, Google DeepMind, Anthropic, and Meta present many of their most important advances without disclosing every technical detail, citing commercial competition, safety, or misuse risks. That practice makes it harder to verify costs, compare methods, and distinguish scientific improvements from product decisions.

R1 shows that at least a substantial part of frontier-model development can be documented to the standard expected by a scientific journal. It also strengthens a line of research with economic implications: if reinforcement learning can draw more reasoning capacity from already pretrained models, companies and academic institutions could experiment without always bearing the cost of building a base system from scratch.

The next step will be to determine how well these results transfer to other models, languages, and tasks that are harder to verify automatically. Nature’s review raises the standard of evidence, but the most demanding test will be whether independent teams can achieve comparable behavior using documented resources and methods.

Turning the headline into a check

An experimental result begins with its observable variable. Record what counts as success, which behavior triggers a label and which cases fall outside scope. review supports publication of the study, not every commercial generalization about the model. If the phenomenon cannot be recognized without interpreting a model's intent, the conclusion needs even greater caution and a reproducible definition.

Protocol matters as much as score. Document instructions, tools, time, compute budget, number of attempts, example selection and grading rule. Changing any one may alter the result without the model learning anything new. Comparing two headlines therefore begins by checking that they measure the same axis.

A strong replication tries to break the conclusion. Add unseen data, small variants, negative controls and tasks where abstention is correct. Preserve failures as well as selected successes. To assess how to separate architecture, stated cost, experimental result and reproducible capability, the set must resemble the intended use and reflect the cost of each error class.

What the record must preserve

A study can reveal a pattern without settling an entire field. Honest wording preserves domain, sample and date, and avoids turning 'we observed' into 'we proved forever.' Evidence becomes more valuable when another team can repeat it with available materials or state what is missing. That traceability is more useful than a sweeping label.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to separate architecture, stated cost, experimental result and reproducible capability. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.