DeepSeek Transparency Report

"2.3.1. Cold Start Unlike DeepSeek-R1-Zero, to prevent the early unstable cold start phase of RL training from the base model, for DeepSeek-R1 we construct and collect a small amount of long CoT data to fine-tune the model as the initial RL actor. To collect such data, we have explored several approaches: using few-shot prompting with a long CoT as an example, directly prompting models to generate detailed answers with reflection and verification, gathering DeepSeek-R1- Zero outputs in a readable format, and refining the results through post-processing by human annotators. In this work, we collect thousands of cold-start data to fine-tune the DeepSeek-V3-Base as the starting point for RL. Compared to DeepSeek-R1-Zero, the advantages of cold start data 9 include: • Readability: A key limitation of DeepSeek-R1-Zero is that its content is often not suitable for reading. Responses may mix multiple languages or lack markdown formatting to highlight answers for users. In contrast, when creating cold-start data for DeepSeek-R1, we design a readable pattern that includes a summary at the end of each response and filters out responses that are not reader-friendly. Here, we define the output format as |special_token||special_token|

, where the reasoning process is the CoT for the query, and the summary is used to summarize the reasoning results. • Potential: By carefully designing the pattern for cold-start data with human priors, we observe better performance against DeepSeek-R1-Zero. We believe the iterative training is a better way for reasoning models."

"2.3.1. Cold Start Unlike DeepSeek-R1-Zero, to prevent the early unstable cold start phase of RL training from the base model, for DeepSeek-R1 we construct and collect a small amount of long CoT data to fine-tune the model as the initial RL actor. To collect such data, we have explored several approaches: using few-shot prompting with a long CoT as an example, directly prompting models to generate detailed answers with reflection and verification, gathering DeepSeek-R1- Zero outputs in a readable format, and refining the results through post-processing by human annotators. In this work, we collect thousands of cold-start data to fine-tune the DeepSeek-V3-Base as the starting point for RL. Compared to DeepSeek-R1-Zero, the advantages of cold start data 9 include: • Readability: A key limitation of DeepSeek-R1-Zero is that its content is often not suitable for reading. Responses may mix multiple languages or lack markdown formatting to highlight answers for users. In contrast, when creating cold-start data for DeepSeek-R1, we design a readable pattern that includes a summary at the end of each response and filters out responses that are not reader-friendly. Here, we define the output format as |special_token||special_token|

, where the reasoning process is the CoT for the query, and the summary is used to summarize the reasoning results. • Potential: By carefully designing the pattern for cold-start data with human priors, we observe better performance against DeepSeek-R1-Zero. We believe the iterative training is a better way for reasoning models."

We exclude sensitive information, credit card numbers, or unique identification information from our training data sources to minimize the risk of collecting any personal information. However, due to the vast scale of pre-training data, some publicly available online content or licensed data from other providers may incidentally contain personal information. We employ technical measures to screen and remove such information from the training data as much as possible and conduct tests before using the data for training. Additionally, to ensure data quality, safety, and diversity, we have established a rigorous data governance process. First, we use filters to automatically screen and remove raw data containing hate speech, pornography, violence, spam, or potential infringement. Second, recognizing that large-scale datasets may inherently contain statistical biases, we combine algorithmic and manual review methods to identify and mitigate the impact of these biases on the model's values, thereby enhancing fairness.

We exclude sensitive information, credit card numbers, or unique identification information from our training data sources to minimize the risk of collecting any personal information. However, due to the vast scale of pre-training data, some publicly available online content or licensed data from other providers may incidentally contain personal information. We employ technical measures to screen and remove such information from the training data as much as possible and conduct tests before using the data for training. Additionally, to ensure data quality, safety, and diversity, we have established a rigorous data governance process. First, we use filters to automatically screen and remove raw data containing hate speech, pornography, violence, spam, or potential infringement. Second, recognizing that large-scale datasets may inherently contain statistical biases, we combine algorithmic and manual review methods to identify and mitigate the impact of these biases on the model's values, thereby enhancing fairness.

From the DeepSeek-V3 technical report: "At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens". Nature paper for R1 notes scale of training data for going from V3 to R1 (e.g. 800k SFT samples of average length of 5k tokens, hence 4B tokens)

The most relevant document is the DeepSeek-V3 paper with the relevant passage on pre-training: "During the pre-training stage, training DeepSeek-V3 on each trillion tokens requires only 180K H800 GPU hours, i.e., 3.7 days on our cluster with 2048 H800 GPUs. Consequently, our pretraining stage is completed in less than two months and costs 2664K GPU hours. Combined with 119K GPU hours for the context length extension and 5K GPU hours for post-training, DeepSeek-V3 costs only 2.788M GPU hours for its full training". Nature paper on R1 notes: For the training of DeepSeek-R1-Zero, we employed 64*8 H800 GPUs, and the process required approximately 198 hours. Additionally, during the training phase of DeepSeek-R1, we utilized the same 64*8 H800 GPUs, completing the process in about 4 days, or roughly 80 hours. To create the SFT datasets, we use 5K GPU hours

The most relevant document is the DeepSeek-V3 paper with the relevant passage on pre-training: "During the pre-training stage, training DeepSeek-V3 on each trillion tokens requires only 180K H800 GPU hours, i.e., 3.7 days on our cluster with 2048 H800 GPUs. Consequently, our pretraining stage is completed in less than two months and costs 2664K GPU hours. Combined with 119K GPU hours for the context length extension and 5K GPU hours for post-training, DeepSeek-V3 costs only 2.788M GPU hours for its full training". No additional information is provided on the R1 phase.

The most relevant document is the DeepSeek-V3 paper with the relevant passage on pre-training: "During the pre-training stage, training DeepSeek-V3 on each trillion tokens requires only 180K H800 GPU hours, i.e., 3.7 days on our cluster with 2048 H800 GPUs. Consequently, our pretraining stage is completed in less than two months and costs 2664K GPU hours. Combined with 119K GPU hours for the context length extension and 5K GPU hours for post-training, DeepSeek-V3 costs only 2.788M GPU hours for its full training". No additional information is provided on the R1 phase.

The most relevant document is the DeepSeek-V3 paper with the relevant passage on compute allocation: "2664K H800 hours on pretraining, 119K on context extension, 5k on post training". Nature paper provides compute allocation of 100k H800 GPU hours for R1-Zero, 5k for SFT data creation, and 41k for R1.

The most relevant documents are the DeepSeek-V3 and DeepSeek-R1 technical reports. The relevant passages state: "During pre-training, we train DeepSeek-V3 on 14.8T high-quality and diverse tokens. The pre-training process is remarkably stable. Throughout the entire training process, we did not encounter any irrecoverable loss spikes or have to roll back. Next, we conduct a two-stage context length extension for DeepSeek-V3. In the first stage, the maximum context length is extended to 32K, and in the second stage, it is further extended to 128K. Following this, we conduct post-training, including Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on the base model of DeepSeek-V3, to align it with human preferences and further unlock its potential" to build V3. Then, "To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates a small amount of cold-start data and a multi-stage training pipeline. Specifically, we begin by collecting thousands of cold-start data to fine-tune the DeepSeek-V3-Base model. Following this, we perform reasoning-oriented RL like DeepSeek-R1- Zero. Upon nearing convergence in the RL process, we create new SFT data through rejection sampling on the RL checkpoint, combined with supervised data from DeepSeek-V3 in domains such as writing, factual QA, and self-cognition, and then retrain the DeepSeek-V3-Base model. After fine-tuning with the new data, the checkpoint undergoes an additional RL process, taking into account prompts from all scenarios. After these steps, we obtained a checkpoint referred to as DeepSeek-R1".

The most relevant documents are the DeepSeek-V3 and DeepSeek-R1 technical reports. The relevant passages state: "During pre-training, we train DeepSeek-V3 on 14.8T high-quality and diverse tokens. The pre-training process is remarkably stable. Throughout the entire training process, we did not encounter any irrecoverable loss spikes or have to roll back. Next, we conduct a two-stage context length extension for DeepSeek-V3. In the first stage, the maximum context length is extended to 32K, and in the second stage, it is further extended to 128K. Following this, we conduct post-training, including Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on the base model of DeepSeek-V3, to align it with human preferences and further unlock its potential" to build V3. Then, "To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates a small amount of cold-start data and a multi-stage training pipeline. Specifically, we begin by collecting thousands of cold-start data to fine-tune the DeepSeek-V3-Base model. Following this, we perform reasoning-oriented RL like DeepSeek-R1- Zero. Upon nearing convergence in the RL process, we create new SFT data through rejection sampling on the RL checkpoint, combined with supervised data from DeepSeek-V3 in domains such as writing, factual QA, and self-cognition, and then retrain the DeepSeek-V3-Base model. After fine-tuning with the new data, the checkpoint undergoes an additional RL process, taking into account prompts from all scenarios. After these steps, we obtained a checkpoint referred to as DeepSeek-R1".

"We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities. Through RL, DeepSeek-R1-Zero naturally emerges with numerous powerful and intriguing reasoning behaviors. However, it encounters challenges such as poor readability, and language mixing." The DeepSeek Hugging Face documentation, notes that the number of parameters associated with the model is 671 billion. "Previous work has heavily relied on large amounts of supervised data to enhance model performance. In this study, we demonstrate that reasoning capabilities can be significantly improved through large-scale reinforcement learning (RL), even without using supervised fine-tuning (SFT) as a cold start. Furthermore, performance can be further enhanced with the inclusion of a small amount of cold-start data. In the following sections, we present: (1) DeepSeek-R1-Zero, which applies RL directly to the base model without any SFT data, and (2) DeepSeek-R1, which applies RL starting from a checkpoint fine-tuned with thousands of long Chain-of-Thought (CoT) examples. 3) Distill the reasoning capability from DeepSeek-R1 to small dense models."

Configuration file released on Hugging Face

"Specifically, we use DeepSeek-V3-Base as the base model and employ GRPO (Shao et al., 2024) as the RL framework to improve model performance in reasoning."

"🔄 DeepSeek-R1 is now MIT licensed for clear open access"

The AI Agent frameworks table in GitHub includes a number of frameworks such as smolagents, YoMo, and superagentx

"We directly apply RL to the base model without relying on supervised fine-tuning (SFT) as a preliminary step. This approach allows the model to explore chain-of-thought (CoT) for solving complex problems, resulting in the development of DeepSeek-R1-Zero. DeepSeekR1-Zero demonstrates capabilities such as self-verification, reflection, and generating long CoTs, marking a significant milestone for the research community. Notably, it is the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT. This breakthrough paves the way for future advancements in this area. We introduce our pipeline to develop DeepSeek-R1. The pipeline incorporates two RL stages aimed at discovering improved reasoning patterns and aligning with human preferences, as well as two SFT stages that serve as the seed for the model’s reasoning and non-reasoning capabilities. We believe the pipeline will benefit the industry by creating better models."

"We evaluate models on MMLU (Hendrycks et al., 2020), MMLU-Redux (Gema et al., 2024), MMLU-Pro (Wang et al., 2024), C-Eval (Huang et al., 2023), and CMMLU (Li et al., 2023), IFEval (Zhou et al., 2023), FRAMES (Krishna et al., 2024), GPQA Diamond (Rein et al., 2023), SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024), SWE-Bench Verified (OpenAI, 2024d), Aider 1 , LiveCodeBench (Jain et al., 2024) (2024-08 – 2025-01), Codeforces 2 , Chinese National High School Mathematics Olympiad (CNMO 2024)3 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024). In addition to standard benchmarks, we also evaluate our models on open-ended generation tasks using LLMs as judges. Specifically, we adhere to the original configurations of AlpacaEval 2.0 (Dubois et al., 2024) and Arena-Hard (Li et al., 2024), which leverage GPT-4-Turbo-1106 as judges for pairwise comparisons. Here, we only feed the final summary to evaluation to avoid the length bias. For distilled models, we report representative results on AIME 2024, MATH-500, GPQA Diamond, Codeforces, and LiveCodeBench."

https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html "Risks associated with AI models may arise from two causes: 1. Limitations due to the immaturity of AI technology. 2. Risks due to the misuse of AI technology. Specifically:" Under "Limitations": - Hallucination Under "Miseuse risks": - Privacy protection - Copyright - Data security - Content safety - Bias - Discrimination https://arxiv.org/pdf/2501.12948 2.3.4. Reinforcement Learning for all Scenarios To further align the model with human preferences, we implement a secondary reinforcement learning stage aimed at improving the model’s helpfulness and harmlessness while simultaneously refining its reasoning capabilities. ... For harmlessness, we evaluate the entire response of the model, including both the reasoning process and the summary, to identify and mitigate any potential risks, biases, or harmful content that may arise during the generation process. Ultimately, the integration of reward signals and diverse data distributions enables us to train a model that excels in reasoning while prioritizing helpfulness and harmlessness. https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html 3.4 You will not use the Services to generate, express or promote content or a chatbot that: (1) is hateful, defamatory, offensive, abusive, tortious or vulgar; (2) is deliberately designed to provoke or antagonize another or is bullying or trolling another; (3) may harass, intimidate, threaten, harm, hurt, scare, distress, embarrass or upset another; (4) is discriminatory such as discriminating another based on race, gender, sexuality, religion, nationality, disability or age; (5) is pornographic, obscene, or sexually explicit (e.g., sexual chatbots); (6) facilitates, promotes, incites or glorifies violence or terrorist/extremism content; (7) exploits, harms, or attempts to exploit or harm or minors or exposes minors to such content; (8) are designed to specifically appeal to or present a persona of any person under the age of 18 or the minimum age required in your country; (9) constitute, encourage or provide instructions for a criminal offence; or (10) impersonates or is designed to impersonate a celebrity, public figure or a person other than yourself without clearly labelling the content or chatbot as "unofficial" or "parody", unless you have that person's explicit consent. 4.4 All Outputs provided by this service are generated by an artificial intelligence model and may contain errors or omissions, for your reference only. You should not treat the Outputs as professional advice. Specifically, when using this service to consult on medical, legal, financial, or other professional issues, please be aware that this service does not constitute any advice or commitment and does not represent the opinions of any professional field. If you require related professional services, you should consult professionals and make decisions under their guidance. The Outputs of this software should not be the basis for your further actions or inactions. Any judgment you make based on the Outputs or subsequent related actions you take will result in consequences and responsibilities borne by you alone, including risks arising from reliance on the truthfulness, accuracy, reliability, non-infringement, or suitability for a specific purpose of the Outputs. You should recognize and use generative artificial intelligence technology scientifically and rationally, and in accordance with the law. 4.5 To prevent confusion and misidentification, and to eliminate or reduce the potential impact of inaccurate generated information, DeepSeek has added prominent disclaimer notices at the end of generated text or at the bottom of interactive interfaces, specifically reminding users that the content is AI-generated. 4.6 DeepSeek can provide incomplete, inaccurate, or offending output. If the output refers to a third-party product or service, this does not mean that the third party approves the comment or is affiliated with DeepSeek. 3.6 You will not use the Services for the following improper purposes: (1) Using the Services for any illegal purposes that violate laws and regulations, regulatory policies, or infringe on the legitimate rights and interests of third parties. (2) Using the Services for dangerous purposes that may have serious harmful impacts on physical health, psychology, society, or the economy, or violate scientific and technological ethics. (3) Engaging in activities that infringe on intellectual property rights, trade secrets, and other violations of business ethics, or using algorithms, data, platforms, etc., to implement monopolistic and unfair competition behaviors. (4) Without DeepSeek's authorization, copying, transferring, leasing, lending, selling, or sub-licensing the entire or part of the Services. (5) Other uses prohibited or restricted by laws and administrative regulations, or that may harm DeepSeek's interests. https://chat.deepseek.com/downloads/DeepSeek%20User%20Agreement.pdf 3.4 DeepSeek advises that, as a user of this Service, you shall not input content that induces outputs violating relevant laws and regulations, including but not limited to: (1) opposing the fundamental principles established by the Constitution; (2) endangering national security, leaking state secrets, subverting state power, overthrowing the socialist system, and undermining national unity; (3) damaging the honor and interests of the state; (4) inciting ethnic hatred and ethnic discrimination and undermining national unity; (5) insulting and abusing the image of martyrdom, denying the deeds of martyrdom, and beautifying and whitewashing the acts of war of aggression; (6) undermining the state's religious policy, promoting cults and feudal superstitions; (7) spreading rumors, false and harmful information, disturbing the social order and undermining social stability; (8) promoting obscenity, pornography, gambling, violence, murder, terrorism, extremism or abetting crime; (9) inciting unlawful assemblies, associations, processions, demonstrations, or gathering of people to disrupt social order; (10) slandering others, revealing their privacy, infringing on their honor, portrait, privacy, intellectual property rights and other legitimate rights and interests; (11) undermining international relations as well as international peace and stability; (12) containing information with other contents prohibited by laws and administrative regulations. 3.5 DeepSeek advises that, as a user of this Service, you shall not input content that induces unfriendly dialogue outputs, including but not limited to: (1) personally attacking and abusing others; (2) making remarks of a cursing, discriminatory, or indifferent to the dignity of life nature against the following groups of people: people of different specific nationalities, geographic regions, genders, gender identities, sexual orientations, races, ethnicities, health conditions, occupations, ages, beliefs, and handicapped groups; (3) cursing, intimidating or threatening others; (4) making direct derogatory comments about content created by other users; (5) using vulgar terms towards other users and creating offense; (6) making comments of a prejudicial nature against the following groups of people, groups that include: specific nationalities, geographic regions, genders, gender identities, sexual orientations, races, ethnicities, health conditions, occupations, ages, beliefs, handicapped groups, etc. 3.6 DeepSeek advises that, as a user of this Service, you shall not maliciously contend with the filtering mechanisms of this Service by, including but not limited to: (1) inputting characters, numbers and other meaningless garbled codes that are difficult to recognize the meaning and affect the reading experience; (2) malicious antagonistic behavior, including but not limited to the use of variants, harmonies and other ways to circumvent the Service detection to enter contents that violates the above 3.4 and 3.5. 3.7 DeepSeek advises that, as a user of this Service, you shall not interfere with the normal operation of this Service or harm DeepSeek's legal rights, including but not limited to: (1) using the Service to engage in illegal and criminal activities such as stealing trade secrets and personal information; (2) reversing compiling, reversing engineer or obtaining any information related to the algorithms of the Service by any other means; (3) using the Service to develop other products and services that compete with the Service (unless the restriction violates relevant legal norms); (4) crawling and copying the Service and any content contained therein by any means (including but not limited to through any robots, crawler technology other automated settings, setting up mirrors); (5) removing or tampering with the artificial intelligence-generated logos or the prominent logos of the deeply synthesized content involved in the Service without the consent of Depth Seeker or without a legally justifiable basis for such removal or tampering; (6) uploading, posting, e-mailing or otherwise transmitting software viruses or other computer codes, files and programs that interfere with, damage or limit the functionality of any computer software, hardware or communication equipment; (7) unlawful intrusion into the servers of the Service, tampering with the code related to the Service, or other acts that damage the Service; (8) attempting to circumvent the security settings or network systems of the DeepSeek, including obtaining data that the user shall not have access to, logging into servers or accounts that are not explicitly authorized, or employing methods such as running port scans to snoop on the security measures of other networks; (9) obtaining output in a manner that causes an unreasonable load to be placed on DeepSeek's computer systems or facilities, or engaging in behavior that could cause such a situation to arise; (10) unlawfully reselling or making available to the public the DeepSeek’s Service or using the DeepSeek’s Service for commercial purposes without authorization from DeepSeek. 3.8 DeepSeek advises that, as a user of this Service, you shall not engage in behaviors that violate other legal regulations or infringe upon the legal rights of others or DeepSeek, such as monopolistic or unfair competition practices.

2.3.4. Reinforcement Learning for all Scenarios To further align the model with human preferences, we implement a secondary reinforcement learning stage aimed at improving the model’s helpfulness and harmlessness while simultaneously refining its reasoning capabilities. ... For harmlessness, we evaluate the entire response of the model, including both the reasoning process and the summary, to identify and mitigate any potential risks, biases, or harmful content that may arise during the generation process. Ultimately, the integration of reward signals and diverse data distributions enables us to train a model that excels in reasoning while prioritizing helpfulness and harmlessness.

Version: 2025-05-28 deepseek-reasoner deepseek-reasoner Model Upgraded to DeepSeek-R1-0528: Enhanced Reasoning Capabilities Significant benchmark improvements (Pass@1) AIME 2025: 70.0 → 87.5 (+17.5) GPQA: 71.5 → 81.0 (+9.5) LCB_v6: 63.5 → 73.3 (+9.8) Aider: 57.0 → 71.6 (+14.6) Note: Complex reasoning tasks may consume more tokens compared to legacy R1 version. Optimized Front-end Development Generated web pages and games now feature improved aesthetics. Reduced Hallucinations Significantly suppressed hallucination issues present in legacy R1 version. JSON Output & Function Calling Support Function call performance: Tau-bench score: 53.5 (Airline) / 63.9 (Retail)

"This is just the beginning! Look forward to multimodal support and other cutting-edge features in the DeepSeek ecosystem."

DeepSeek API: https://api-docs.deepseek.com/guides/reasoning_model ModelScope: https://huggingface.co/deepseek-ai/DeepSeek-R1-0528 HuggingFace: https://www.modelscope.cn/models/deepseek-ai/DeepSeek-R1-0528/summary

Model weights quantization example config: ``` "quantization_config": { "activation_scheme": "dynamic", "fmt": "e4m3", "quant_method": "fp8", "weight_block_size": [ 128, 128 ] } ``` API quantization: "All DeepSeek-V3/R1 inference services are served on H800 GPUs with precision consistent with training. Specifically, matrix multiplications and dispatch transmissions adopt the FP8 format aligned with training, while core MLA computations and combine transmissions use the BF16 format, ensuring optimal service performance."

This code repository and the model weights are licensed under the MIT License. DeepSeek-R1 series support commercial use, allow for any modifications and derivative works, including, but not limited to, distillation for training other LLMs. Please note that: https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html

For harmlessness, we evaluate the entire response of the model, including both the reasoning process and the summary, to identify and mitigate any potential risks, biases, or harmful content that may arise during the generation process. Ultimately, the integration of reward signals and diverse data distributions enables us to train a model that excels in reasoning while prioritizing helpfulness and harmlessness.

" {"role": "system", "content": "You are a helpful assistant."},"

R1's output includes reasoning content

You represent and warrant that Services may not be used in or for the benefit of, or exported, re-exported, or transferred (a) to or within any country subject to comprehensive sanctions under Export Control and Sanctions Laws; (b) to any party on any restricted party lists under any applicable Export Control and Sanctions Laws that would prohibit your use of Services.

R1 released on HF under an MIT license with no prohibited uses. Section 3.4 of terms of use: ""3.4 You will not use the Services to generate, express or promote content or a chatbot that:...""

For users who violate regulations, DeepSeek has the right, according to its reasonable judgment and without notice, to take actions such as warnings, functionality restrictions, service suspensions or terminations, account bans, content deletions, and to save related records and report to competent authorities

We also participated in the China AI Safety and Security Commitments Framework. Moreover, China’s AI regulatory framework differs significantly from that of the U.S., relying on formal legislation and regulatory oversight rather than solely on voluntary corporate commitments. As such, participation in commitments should not be used as a direct basis for comparison.