DeepSeek-R1

DeepSeek-R1 is an open-weights reasoning model released by DeepSeek on 22 January 2025. It is built on the DeepSeek-V3 base and was published as arXiv 2501.12948 and later in Nature 645, 633–638 (2025). The model demonstrated frontier-competitive reasoning at lower training compute than leading US models and became a reference case in debates over fast-follow competition, distillation, and open-weights release.

Capabilities and benchmarks

In February 2025 DeepSeek-R1 briefly matched the top US model, OpenAI's o1, an event an Epoch AI account described as "driving the stock market bananas" (Source: epochai.substack.com). It showed that compute-poor labs could reach frontier-competitive reasoning performance using less training compute.

The R1 paper reports that DeepSeek-R1 performs comparably to OpenAI-o1-1217 on reasoning tasks (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning). In mathematics it scored 79.8% pass@1 on AIME 2024 against 79.2% for o1-1217, 97.3% on MATH-500 against 96.4%, and 78.8% on CNMO 2024. In competitive programming it scored 65.9% on LiveCodeBench and a Codeforces rating of 2,029, above 96.3% of human participants, against 2,061 for o1-1217. On knowledge benchmarks it scored 90.8% on MMLU, 84.0% on MMLU-Pro and 71.5% on GPQA Diamond, above DeepSeek-V3 and below o1-1217 on MMLU and GPQA Diamond, and 30.1% on SimpleQA. On engineering-oriented coding it scored 49.2% on SWE-bench Verified (o1-1217: 48.9%) and 53.3% on Aider-Polyglot (61.7%). It reached a length-controlled win rate of 87.6% on AlpacaEval 2.0 and 92.3% on Arena-Hard. DeepSeek took the o1-1217 figures from OpenAI's official reports because the o1-1217 API was difficult to access from mainland China (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning).

The paper lists four limitations. DeepSeek-R1 falls short of DeepSeek-V3 on function calling, multi-turn dialogue, complex role-playing and JSON output; it is optimized for Chinese and English and may reason and answer in English when queried in other languages; few-shot prompting consistently degrades its performance, so DeepSeek recommends zero-shot prompts that describe the problem and the output format; and because long evaluation times limited large-scale RL on software-engineering tasks, it shows little improvement over DeepSeek-V3 on those benchmarks (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning).

Training and architecture

The R1 paper (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning) introduces two training regimes:

• R1-Zero — pure reinforcement learning (RL) from a base model with no supervised fine-tuning (SFT). Reasoning behaviors such as self-reflection, verification, and long chain-of-thought emerge from RL with rule-based rewards. The paper's central claim is that reasoning does not require human-labeled reasoning trajectories.
• R1 — adds a cold-start SFT stage and multi-stage RL to fix the readability and language-mixing issues seen in R1-Zero.

DeepSeek-R1 has the same Mixture-of-Experts architecture as DeepSeek-V3, with 671 billion total and 37 billion activated parameters. R1-Zero was trained from DeepSeek-V3-Base with Group Relative Policy Optimization (GRPO), which estimates the baseline from the scores of a group of sampled outputs and so needs no critic model. Its rewards were rule-based: an accuracy reward (a boxed final answer checked by rule for math, compiled code run against test cases for programming problems) and a format reward for placing reasoning between <think> tags. DeepSeek did not use a neural reward model, citing the risk of reward hacking in large-scale RL. Over thousands of RL steps R1-Zero's average AIME 2024 pass@1 rose from 15.6% to 71.0%, and 86.7% with majority voting over 64 samples; its responses grew from hundreds to thousands of reasoning tokens, and the paper describes an "aha moment" in which an intermediate checkpoint learned to stop and re-evaluate its approach. Its outputs were hard to read and mixed languages (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning).

R1's pipeline has four stages (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning):

1. Cold start. DeepSeek-V3-Base is fine-tuned on thousands of long chain-of-thought examples, gathered by few-shot and direct prompting, from readable R1-Zero outputs, and through post-processing by human annotators, in a format that ends each response with a summary.
2. Reasoning-oriented RL. The same large-scale RL as R1-Zero, on coding, math, science and logic, with an added language-consistency reward equal to the proportion of target-language words in the chain of thought. DeepSeek reports that this reward slightly lowered benchmark performance but made the output more readable.
3. Rejection sampling and SFT. About 600,000 reasoning samples, kept only when correct and filtered for mixed languages, long paragraphs and code blocks, are combined with about 200,000 non-reasoning samples (writing, factual QA, self-cognition, translation) from the DeepSeek-V3 pipeline; DeepSeek-V3-Base is fine-tuned on the roughly 800,000 samples for two epochs.
4. RL for all scenarios. Rule-based rewards for reasoning plus preference reward models for general data, with helpfulness judged on the final summary and harmlessness on the whole response, reasoning included.

DeepSeek also reports approaches that did not work in its early experiments: process reward models, limited by the difficulty of defining and grading intermediate steps and by reward hacking, and Monte Carlo tree search, limited by the size of the token search space and the difficulty of training a fine-grained value model. In its comparison of distillation and RL on a 32B model, the paper concludes that distillation is economical and effective but that advancing further may still require more powerful base models and larger-scale RL, a conclusion relevant to Scaling Laws (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning).

Safety

The final RL stage evaluated harmlessness over the whole response, including the reasoning, to identify risks, biases or harmful content. The paper attributes DeepSeek-R1's lower score than DeepSeek-V3 on Chinese SimpleQA (63.7% against 68.0%) to refusals introduced by safety RL, and states that without safety RL its accuracy would exceed 70% (DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning).

Publication and peer review

The R1 paper appeared on the cover of Nature on September 18, 2025 (vol. 645, issue 8081, pp. 633–638), following formal peer review — coverage of the publication described R1 as the first mainstream large language model to pass through journal peer review (Source: nature.com; eu.36kr.com). Supplementary material published with the paper disclosed that the R1 reasoning-stage training cost was $294,000 in compute (on top of the base-model cost), a figure widely cited as far below the reported training costs of comparable US models (Source: mlq.ai).

Open weights and distillation

DeepSeek-R1 was released publicly with open weights, enabling distillation by other labs (Source: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning). Dense-model distillations shipped in 1.5B, 7B, 8B, 14B, 32B, and 70B sizes (Source: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning). They were made by fine-tuning Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, Qwen2.5-14B, Qwen2.5-32B, Llama-3.1-8B and Llama-3.3-70B-Instruct on the roughly 800,000 samples curated with DeepSeek-R1, using SFT only. DeepSeek-R1-Distill-Qwen-32B scored 72.6% on AIME 2024, 94.3% on MATH-500 and 57.2% on LiveCodeBench, results the paper describes as comparable to OpenAI's o1-mini; the 32B and 70B distilled models exceeded o1-mini on most of the paper's reasoning benchmarks, the 14B model surpassed QwQ-32B-Preview on all of them, and the 7B model outperformed GPT-4o-0513 on all of them. Applying more than 10,000 steps of large-scale RL directly to Qwen-32B-Base produced a model on par with QwQ-32B-Preview and below the distilled 32B model on every benchmark, from which the paper concludes that distilling a stronger model works better for small models than RL on them alone (Source: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning). Distillations into Qwen2.5 showed efficiency gains the Epoch AI account called "enormous" — several-fold less compute for GPQA and multiple orders of magnitude less for MATH (Source: epochai.substack.com). The model was released with roughly 4B tokens of distillation data, a relatively small amount but sufficient to meaningfully improve post-training (Source: epochai.substack.com). Anthropic accused DeepSeek of distilling from Claude's outputs (Source: epochai.substack.com).

The R1-to-small-dense distillations serve as a reference demonstration of reasoning distillability (see Distillation).

Reception and competitive context

DeepSeek-R1 caught up to OpenAI's o1 within months using less training compute, a case cited in discussions of the fast-follow problem. One account notes that if software efficiency improves on the order of 10× per year, the speed of the catch-up is less surprising (Source: The Least Understood Driver of AI Progress; see AI Software Progress). The open release prompted other Chinese labs to follow suit, including Baidu abandoning a closed-model commitment (see Open-Source AI / Open-Weight Models).

Related models

• DeepSeek-V3 — the base model R1 was built on.
• DeepSeek V4 Pro / V4 Flash — DeepSeek's later flagship generation.
• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — the model's technical paper.
• State of AI Report 2025 — covers R1 in the year's competitive landscape.

Relationships

• instance-of: Frontier Models, Reasoning Models and Chain-of-Thought
• depends-on: DeepSeek-V3 — base model
• related: AIME (as a model benchmark), MATH and MATH-500, GPQA, LiveCodeBench, SWE-bench — benchmarks carrying the R1 paper's results for R1, R1-Zero and the distilled models
• supports: Fast-Follow Problem — cited as the reference fast-follow case
• related: Distillation, Open-Source AI / Open-Weight Models, AI Software Progress, US-China AI Competition: Different Races, Different Metrics