DeepSeek-R1

DeepSeek-R1 is an open-weight reasoning model and large language model family developed by the Chinese artificial intelligence laboratory DeepSeek. The original model was released on January 20, 2025. It uses the architecture of DeepSeek-V3-Base and a post-training pipeline designed to produce extended reasoning traces before a final answer. DeepSeek released the R1 and R1-Zero weights, six smaller distilled models, and a technical report.^[1][2]

The original R1 checkpoint has 671 billion total parameters, with 37 billion activated for each token, and a 128,000-token context window. Its base is a sparse mixture-of-experts model rather than a 671-billion-parameter dense model. DeepSeek-R1-Zero was trained from the base checkpoint using reinforcement learning without preliminary supervised fine-tuning. The final DeepSeek-R1 added cold-start examples, two reinforcement-learning stages, rejection sampling, and supervised training on both reasoning and non-reasoning data.^[2][3]

DeepSeek's release materials called R1 "open source," but open-weight is the more precise description. The trained weights and repository are under the MIT License, and the published work provides data samples, aggregate data statistics, an inference script, and extensive methodological details. It does not release the complete pre-training corpus, the complete post-training datasets, or DeepSeek's internal distributed training framework.^[2][3]

Scroll sideways for more →

Release and model family

DeepSeek previewed its reasoning-model work as DeepSeek-R1-Lite-Preview in November 2024. The preview was available only through the company's website. Its announcement did not disclose a parameter count or base checkpoint, and the preview did not yet have released weights or API access. DeepSeek said open-source models and an API were coming.^[4]

The January 2025 release contained three related forms rather than one interchangeable model:

Scroll sideways for more →

R1-Zero was a research checkpoint intended to test whether a large pre-trained model could improve on verifiable reasoning tasks through reinforcement learning alone. Its authors reported substantial gains, but also poor readability, mixed Chinese and English within some traces, and weak performance outside the reasoning domains emphasized by its rewards. The final R1 pipeline was designed to retain the reasoning gains while improving presentation, general instruction following, helpfulness, and safety.^[3]

The R1 release also exposed a generated reasoning trace before the final answer. In the original API, this appeared separately from the final response, and local checkpoints used <think> and </think> delimiters. The trace is observable generated text. It should not be equated with human thought or assumed to be a complete causal explanation of the model's answer.^[2][5][6]

Architecture

R1 and R1-Zero use the Transformer and expert-routing design of DeepSeek-V3-Base. The V3 report specifies 61 Transformer layers and a hidden dimension of 7,168. Its Multi-head Latent Attention configuration uses 128 attention heads and compresses key-value representations into a lower-dimensional latent representation. This design is intended to reduce the memory required for the attention cache during inference.^[7]

Most feed-forward layers are sparse expert layers. Each such layer contains one shared expert and 256 routed experts, with eight routed experts selected for each token. The resulting checkpoint contains 671 billion parameters, but only about 37 billion are active for a token. Parameter count alone therefore does not describe its inference cost in the same way as it does for a dense model.^[7]

DeepSeek-V3 also used a one-step multi-token-prediction objective during pre-training. Its context window was extended in two stages, first from 4,000 to 32,000 tokens and then to 128,000 tokens. R1 retained the 128,000-token open-checkpoint context length. The original R1 evaluation capped generated answers at 32,768 tokens even though the input context could be longer.^[2][3][7]

The architecture is not the main difference between V3-Base, R1-Zero, and R1. They share the same large base design. Their important differences are the post-training data, reward functions, and sequence of supervised and reinforcement-learning stages.

Training

Group Relative Policy Optimization

DeepSeek trained the R1 models with Group Relative Policy Optimization, an algorithm first described in the DeepSeekMath report. For each prompt, GRPO samples a group of candidate outputs, scores them, and calculates each output's advantage relative to the reward distribution within that group. It does not train a separate value model comparable in size to the policy model. The original GRPO paper presented this as a way to reduce the memory and computation required by value-model-based policy optimization.^[8]

The R1 implementation included a reference policy and a Kullback-Leibler regularization term. In R1-Zero training, each question produced 16 sampled outputs. The peer-reviewed methods report a sampling temperature of 1, a batch of 32 unique questions per step, and a maximum generated length of 32,768 tokens before step 8,200 and 65,536 tokens afterward. Training continued for 10,400 policy-update steps, corresponding to about 1.6 epochs over the prompt set.^[3]

R1-Zero rewards

R1-Zero began directly from DeepSeek-V3-Base and used reinforcement learning without a preliminary supervised stage. Its rule-based reward combined:

• an accuracy reward for domains with mechanically checkable outcomes, such as a mathematical answer or code evaluated against tests; and
• a format reward for placing the generated reasoning and final answer in the requested tags.

The authors deliberately avoided a learned neural reward model for R1-Zero's reasoning tasks. They argued that a learned reward could be exploited during large-scale reinforcement learning, while deterministic checkers supplied clearer signals in mathematics, code, and formal logic. This choice also limited the practical training distribution to tasks for which a reliable checker could be constructed.^[3]

During training, the reported AIME 2024 average pass@1 score rose from 15.6% to 77.9%, while majority-vote consistency over 16 samples reached 86.7%. Average output length also increased substantially. These are results reported by the model's developers for one benchmark and training run, not evidence that longer reasoning always causes a correct answer.^[3]

The final R1 pipeline

The final R1 training process used four broad stages:

1. Cold start: DeepSeek collected thousands of examples with a more readable, human-oriented reasoning format and used them to supervise the base model before reinforcement learning.
2. Reasoning reinforcement learning: The model was optimized on reasoning prompts using rule-based rewards. A language-consistency reward was added to reduce unintended language mixing.
3. Rejection sampling and supervised fine-tuning: Outputs from the reasoning checkpoint were sampled and filtered. These were combined with non-reasoning examples for another supervised fine-tuning stage.
4. Mixed-objective reinforcement learning: A final stage combined rule-based rewards for reasoning data with model-based preference and safety rewards for general prompts.^[3]

The supervised dataset in stage three contained about 800,000 examples: approximately 600,000 reasoning samples and 200,000 non-reasoning samples. For reasoning data, DeepSeek sampled several trajectories per prompt and retained correct outputs. Some candidates were judged with DeepSeek-V3 when a deterministic rule was unavailable. The authors say they filtered mixed-language traces, long unstructured paragraphs, and code blocks from this dataset. Non-reasoning data covered areas such as writing, factual question answering, translation, software engineering, and front-end development.^[9]

For the final preference-oriented stage, DeepSeek reports curating 66,000 response pairs for a helpfulness reward model. The helpfulness evaluator was trained to judge the final summary while minimizing interference with the reasoning trace. A separate safety reward model used 106,000 prompts with model-generated responses labeled safe or unsafe. The authors evaluated the entire response, including the reasoning trace, for safety.^[3]

These details qualify the often repeated description of R1 as trained with "pure RL." That description applies to R1-Zero's route from V3-Base, not to the final R1 model. Final R1 used supervised cold-start data, a large supervised dataset constructed after rejection sampling, and model-based rewards in addition to rule-based reinforcement learning.^[3][9]

Generated reasoning and the "aha moment"

DeepSeek's paper used the phrase "aha moment" for a late-training increase in reflective language, especially the token "wait," in one R1-Zero example. The supplement provides a broader analysis: reflective expressions increased gradually overall, while some specific expressions became more common at particular training steps.^[3][9]

The label should not be read as proof of a discrete internal cognitive event. A Sea AI Lab pilot study of smaller R1-Zero-like runs found reflection-like expressions in base models before reinforcement learning and found that longer responses could result from optimizing the reward rather than from a newly acquired self-reflection faculty. That study did not reproduce DeepSeek's full 671-billion-parameter run, so it narrows the interpretation of the "aha" narrative rather than disproving the reported R1-Zero training curve.^[10]

Compute and cost accounting

The most important cost distinction is between R1-family post-training, DeepSeek-V3 base-model training, and a complete research-and-development budget. They are not the same quantity.

DeepSeek's peer-reviewed supplement reports the following R1-family compute. Dollar figures are rental-equivalent estimates that assume a price of $2 per H800 GPU-hour.^[9]

Scroll sideways for more →

The supplement says R1-Zero used 512 Nvidia H800 GPUs for approximately 198 hours. The final R1 reinforcement-learning training used the same number of H800 GPUs for roughly 80 hours, which corresponds to about 41,000 GPU-hours. It also mentions preparatory experiments on a smaller 30-billion-parameter model using A100 GPUs, but does not assign those experiments a cost in the table.^[9]

Accordingly, $294,000 is not the cost of the final R1 reinforcement-learning stage. It is the table's total for R1-Zero reinforcement learning, supervised-data creation, and final R1 reinforcement learning. The final R1 row itself is $82,000. The estimate is not a cash-spending audit and does not include every research experiment, staff cost, data-acquisition cost, inference expense, or the cost of training the V3 base checkpoint.^[9]

DeepSeek separately estimated the official DeepSeek-V3 training run at 2.788 million H800 GPU-hours, or $5.576 million at the same assumed $2-per-hour rate. That number includes V3 pre-training, context extension, and post-training. The V3 report expressly excludes earlier architecture research, algorithm experiments, and ablation studies. It is therefore inaccurate both to call $5.576 million the cost of R1 and to present $294,000 as the complete cost of developing the R1 system.^[7]

Evaluation

DeepSeek reported strong results in competition mathematics, science question answering, algorithmic coding, and some instruction-following evaluations. Cross-model comparisons require care because model versions, prompts, tools, token limits, benchmark snapshots, and sampling procedures differ.

For the peer-reviewed evaluation, generated answers were capped at 32,768 tokens. The authors used temperature 0.6 and top-p 0.95 rather than greedy decoding. They generated between four and 64 responses per question, depending on the benchmark, and estimated pass@1 as the mean correctness of those samples. They used 64 samples for AIME and GPQA, 16 for MATH and Codeforces, and eight for LiveCodeBench. This is different from majority-vote consensus or best-of-N scoring.^[9]

Selected developer-reported results for the original R1 are shown below. They reproduce the metric and snapshot used in the Nature publication rather than combining later benchmark versions.^[3][9]

Scroll sideways for more →

The paper also reported contamination controls. DeepSeek filtered pre-training and post-training material when it contained matching 10-token sequences from evaluation questions or reference solutions, and says its mathematical SFT and RL prompts came from competitions before 2023. The authors acknowledged that this method cannot detect paraphrased test material, especially for benchmarks released before 2024.^[9]

The January release said R1 performed "on par" with OpenAI o1 on several reasoning tasks. The paper's tables did report similar values on some selected benchmarks, but not a uniform advantage. For example, the reported R1 value was higher on the paper's LiveCodeBench snapshot, while o1-1217 was higher on GPQA Diamond and Aider-Polyglot. Some o1 values were taken from OpenAI's own reports because the API was difficult for the DeepSeek authors to access in mainland China. These results support benchmark-specific comparisons, not a claim that either model was universally better.^[1][9]

Evaluation numbers should also be separated from real-world reliability. Competition tasks often have short, verifiable answers and match the reward design used for R1's reasoning training. High scores on those tasks do not establish accuracy on arbitrary factual, legal, medical, or open-ended questions. The Nature paper itself reported only limited gains from reasoning-specific RL on general preference benchmarks until non-reasoning supervised data and mixed rewards were added.^[3]

Distilled models

Knowledge distillation was used to transfer patterns from the large R1 model into smaller dense checkpoints. DeepSeek fine-tuned four Qwen 2.5 bases and two Llama bases on approximately 800,000 R1-curated examples.^[2][9]

Scroll sideways for more →

The distilled checkpoints are not compressed copies of the 671-billion-parameter network. They retain the architecture and much of the prior knowledge of their Qwen or Llama base and are fine-tuned to imitate selected R1 outputs. They therefore differ from full R1 in capacity, behavior, deployment requirements, and license obligations.^[2]

DeepSeek reported that its 32B Qwen distill reached 72.6% pass@1 on AIME 2024, 94.3% on MATH-500, and 57.2% on LiveCodeBench under the release evaluation. The 70B Llama distill was reported at 70.0%, 94.5%, and 57.5%, respectively. These are developer evaluations under the stated sampling setup, not independent certification.^[2]

Availability and licensing

The R1 repository, R1-Zero weights, R1 weights, and DeepSeek-authored code are distributed under the MIT License. The license permits use, modification, redistribution, and commercial use subject to its notice and disclaimer. DeepSeek's release also allowed outputs from its hosted R1 API to be used for fine-tuning and distillation.^[1][2]

At launch, DeepSeek exposed R1 through the deepseek-reasoner API name. Its published introductory prices were $0.14 per million cached input tokens, $0.55 per million uncached input tokens, and $2.19 per million output tokens. Those were January 2025 service prices, not model-training costs, and they should not be presented as current pricing after the hosted alias moved to later models.^[1][11]

The distilled models require an additional distinction. The Qwen-based distills derive from models released under Apache 2.0, while the Llama-based distills remain subject to the corresponding Llama 3.1 or Llama 3.3 license. The R1 repository instructs users to account for those base-model licenses. "MIT licensed" should therefore not be used to imply that every dependency and every distilled checkpoint has only MIT terms.^[2]

DeepSeek's Nature code-availability statement lists public R1 and R1-Zero weights, a V3 inference script, and common libraries including PyTorch and vLLM. It also says the distributed training system was based on DeepSeek's internal HAI-LLM framework. The data-availability statement provides samples used for rejection sampling and RL prompts, together with statistics and methodology, but not the full datasets or V3 pre-training corpus.^[3]

These facts make the release unusually inspectable for a model of its scale, while still falling short of a fully reproducible build from raw data and public training infrastructure. Independent projects such as Hugging Face's Open-R1 were started specifically to reproduce missing pieces of the data and training pipeline in public.^[12]

DeepSeek-R1-0528

DeepSeek released DeepSeek-R1-0528 on May 28, 2025. The company said it kept the same V3-Base foundation but used more post-training compute and algorithmic changes. It added support for system prompts, JSON output, and function calling.^[13][14]

The 0528 release used two parameter-count conventions that require explanation. Original R1 materials report 671 billion total parameters and 37 billion active parameters. The 0528 release page reports 685 billion parameters when a 14-billion-parameter multi-token-prediction layer is included. Because DeepSeek says 0528 uses the same base architecture, these disclosures indicate that the difference reflects whether that auxiliary layer is included in the headline total rather than the addition of a new 14-billion-parameter reasoning backbone.^[2][7][13]

DeepSeek reported a 128,000-token context length for the downloadable 0528 checkpoint. Its own website, mobile application, and API exposed a 64,000-token context at the time of the update, with a maximum combined reasoning-and-answer output of 64,000 tokens.^[13]

Selected 0528 results from DeepSeek's model card are shown below. Both columns are developer-reported. The LiveCodeBench row uses an August 2024 to May 2025 snapshot, so its 63.5 baseline should not be combined with the 65.9 result from the earlier January-ending snapshot in the Nature paper.^[14]

Scroll sideways for more →

The label used by DeepSeek's hosted API later stopped identifying a stable R1 checkpoint. At release, deepseek-reasoner referred to R1, and in May 2025 it was updated to R1-0528. DeepSeek subsequently routed the same alias to later V3-series thinking modes. In April 2026, the company announced that deepseek-chat and deepseek-reasoner would be retired after July 24, 2026; during the transition, both aliases routed to modes of DeepSeek-V4-Flash. The downloadable R1 and R1-0528 weights remained separate artifacts regardless of the hosted alias.^[1][13][11]

Limitations and research scrutiny

Reasoning length and overthinking

The R1 authors found that output length grew during R1-Zero training and that difficult prompts often elicited longer traces. They also documented excessive reasoning on simple questions, sensitivity to prompt wording, weak structured-output behavior in the original version, language mixing outside Chinese and English, and limited tool use. They described software-engineering RL as constrained by the time required to evaluate generated patches.^[3]

A 2026 Transactions on Machine Learning Research study examined R1 across reasoning length, long context, culture, safety, and other controlled settings. It found a task-dependent "sweet spot" beyond which extra generated reasoning could reduce performance, and described repeated reconstruction of already explored solutions as rumination. These findings concern the study's tasks and inference settings; they do not imply that every long R1 response is worse than a short one.^[5]

Faithfulness of the visible trace

A visible chain of thought is not guaranteed to report every factor that influenced the answer. Chen and colleagues inserted six kinds of hints into multiple-choice prompts and measured whether a model acknowledged a hint after its answer changed toward the hinted option. Averaged across their settings, DeepSeek-R1 verbalized the hint in 39% of qualifying cases.^[6]

The experiment has important limits. It used benchmark-style multiple-choice questions and artificial hints that were easy for a model to exploit, including settings where the intended behavior did not require extended reasoning. The authors presented the result as evidence against a guarantee of faithful traces, not as proof that 61% of every R1 trace is false or useless.^[6]

Safety and jailbreaks

DeepSeek's own evaluation described the unmodified model's safety as moderate and reported improved results when it was paired with a separate risk-control system. The paper also warned that R1 could be jailbroken into producing dangerous content and that open weights could be further fine-tuned in ways that weaken safeguards. Its supplement evaluated standard safety datasets, 50 languages, and a set of 2,232 jailbreak instructions, but many labels were produced by another language model acting as judge.^[3][9]

The Thoughtology study independently reported greater vulnerability for R1 than for V3 in its tested harmful-generation setting and found that R1-generated attacks could elicit harmful responses from other safety-aligned models. As with the vendor evaluation, these results depend on the prompts, attack method, target systems, and safety classifier used in the study.^[5]

Results for the downloadable 671-billion-parameter weights, a smaller distilled checkpoint, a third-party host, and DeepSeek's own website or API are not interchangeable. A host can add system prompts, input filters, output filters, retrieval, or other controls that are absent from the raw checkpoint. Conversely, a local deployment can change or remove those layers.

Politically sensitive responses

The R1dacted study tested what it described as a locally deployed full R1 checkpoint and reported refusal or template-like responses on politically sensitive China-related prompts. Its result indicates that at least some such behavior could persist outside DeepSeek's hosted service rather than being solely a website-level output filter.^[15]

Its percentages need careful interpretation. The study's main 10,030-prompt dataset was selected through a pipeline specifically designed to retain English prompts that triggered its definition of local censorship in R1 while excluding prompts that other tested models also refused on general safety grounds. It therefore does not estimate a censorship rate over arbitrary user questions. In a separate balanced test of 2,880 generated prompts across 96 preselected sensitive categories, the study reported very high refusal rates in many categories.^[15]

The same paper found only limited transfer of its two primary refusal patterns to the six distilled models. On the selected 10,030 prompts, the reported total rates for those models ranged from 0.15% to 0.30%. This difference is another reason not to treat results from an R1 distill as results from the full R1 checkpoint.^[15]

Reproducibility

The Nature publication substantially expanded the details available in the January preprint, including training hyperparameters, data-generation statistics, reward-model datasets, compute accounting, decontamination procedures, and safety tests. Important gaps remain: DeepSeek did not release the complete training data, the internal HAI-LLM framework, or everything required to reproduce the 671-billion-parameter run exactly.^[3][9]

Public projects explored smaller-scale reproductions. Open-R1 organized open datasets, evaluation code, supervised fine-tuning, and reinforcement-learning recipes. TinyZero reproduced a limited R1-Zero-like setup on countdown and multiplication tasks with a 3-billion-parameter base model. These projects demonstrate transferable techniques, but neither is a reproduction of the full R1 model, data distribution, or compute scale.^[12][16]

Reception and impact

R1 attracted widespread attention outside the research community in late January 2025. The Associated Press reported that DeepSeek's assistant became the most-downloaded free iPhone application in the United States and that Nvidia's share price fell 17% on January 27 amid investor concern that competitive AI models might require less cutting-edge hardware or capital than expected. Nvidia simultaneously described DeepSeek's work as an AI advance.^[17]

That market reaction did not verify a total development cost. Early coverage often compared the $5.576 million V3 training estimate, the $294,000 R1-family post-training estimate, and much larger corporate infrastructure budgets as though they measured the same thing. The published reports show that they do not. R1 still depended on a large V3 base model, hundreds of H800 accelerators for post-training, and substantial inference compute for long sampled responses.^[9][7][17]

Within research, the release accelerated work on reinforcement learning from verifiable rewards, smaller distilled reasoning models, evaluation of visible reasoning traces, and public replication projects. Its influence came from the combination of capable weights, a relatively detailed training account, and permission to modify and distill outputs, not from proof that reasoning models can be built without substantial data or compute.^{[2][3][12][16]}

Researchers interviewed by Nature soon after the release emphasized the opportunity to inspect and adapt a capable reasoning model whose weights were available. They also noted that the training data remained undisclosed and that running the full model locally still required substantial hardware.^[18]

Peer-reviewed publication

The R1 work first appeared as an arXiv preprint in January 2025. A substantially expanded version was received by Nature on February 14, accepted on July 17, and published online on September 17, 2025. The peer-reviewed article and its 83-page supplement added methodological, compute, data, evaluation, and safety details that were not available in the initial release.^[3][9]

See also

• Reinforcement learning from verifiable rewards
• Test-time compute

References

1. ^DeepSeek. "DeepSeek-R1 Release." January 20, 2025. api-docs.deepseek.com/...news250120
2. ^DeepSeek-AI. "DeepSeek-R1" repository and model card. 2025. github.com/...DeepSeek-R1
3. ^Guo, D., Yang, D., Zhang, H., et al. "DeepSeek-R1 Incentivizes Reasoning in LLMs Through Reinforcement Learning." Nature 645, 633-638. September 17, 2025. doi.org/...s41586-025-09422-z
4. ^DeepSeek. "DeepSeek-R1-Lite-Preview Is Now Live." November 20, 2024. api-docs.deepseek.com/...news1120
5. ^Marjanovic, S. V., Patel, A., Adlakha, V., et al. "DeepSeek-R1 Thoughtology: Let's Think About LLM Reasoning." Transactions on Machine Learning Research. January 2026. openreview.net/forum
6. ^Chen, Y., Benton, J., Radhakrishnan, A., et al. "Reasoning Models Don't Always Say What They Think." 2025. arxiv.org/...2505.05410
7. ^DeepSeek-AI. "DeepSeek-V3 Technical Report." 2024. arxiv.org/...2412.19437
8. ^Shao, Z., Wang, P., Zhu, Q., et al. "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models." 2024. arxiv.org/...2402.03300
9. ^Guo, D., Yang, D., Zhang, H., et al. "Supplementary Information for DeepSeek-R1 Incentivizes Reasoning in LLMs Through Reinforcement Learning." 2025. media.springernature.com/...25_9422_MOESM1_ESM.pdf
10. ^Liu, Z., Chen, C., Li, W., Pang, T., Du, C., and Lin, M. "There May Not Be Aha Moment in R1-Zero-Like Training: A Pilot Study." Sea AI Lab. February 7, 2025. sail.sea.com/...62
11. ^DeepSeek. "DeepSeek-V4 Preview Release." April 24, 2026. api-docs.deepseek.com/...news260424
12. ^Hugging Face. "Open-R1: A Fully Open Reproduction of DeepSeek-R1." January 28, 2025. huggingface.co/...open-r1
13. ^DeepSeek. "DeepSeek-R1-0528 Release." May 28, 2025. api-docs.deepseek.com/...news250528
14. ^DeepSeek-AI. "DeepSeek-R1-0528" model card. May 2025. huggingface.co/...DeepSeek-R1-0528
15. ^Naseh, A., Chaudhari, H., Roh, J., et al. "R1dacted: Investigating Local Censorship in DeepSeek's R1 Language Model." 2025. arxiv.org/...2505.12625
16. ^Pan, J., Zhang, J., Wang, X., Yuan, L., Peng, H., and Suhr, A. "TinyZero: Minimal Reproduction of DeepSeek R1-Zero." 2025. github.com/...TinyZero
17. ^O'Brien, M. and Chan, K. "What Is DeepSeek, the Chinese AI Company Upending the Stock Market?" Associated Press. January 27, 2025. apnews.com/...f4908eaca221d601e31e7e3368778030
18. ^Gibney, E. "China's Cheap, Open AI Model DeepSeek Thrills Scientists." Nature 638, 13-14. January 23, 2025. doi.org/...d41586-025-00229-6