DeepSeek-R1
In plain wordsA free, downloadable AI reasoning model released by the Chinese company DeepSeek in January 2025. It matched leading American models at math and coding for a fraction of the reported cost, shaking the industry.
An open-weightopen-weight model, n.: An AI model whose trained internal numbers are published for anyone to download and run on their own computers, even if the data and code used to make it stay private. reasoning modelreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning. released by DeepSeekDeepSeek, prop. n.: A Chinese artificial intelligence company whose free, cheaply trained models startled the world in January 2025 by nearly matching the best American systems. on 20 January 2025 under the MIT license. Built on the DeepSeek-V3 base model, a mixture-of-expertsmixture of experts, n.: A way to build very large AI models cheaply: the model contains many specialist sub-networks, and for each word only a few are switched on, picked by a small dispatcher network. transformertransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it. with 671 billion parameters of which about 37 billion are active for each token, it was trained with large-scale reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds. using group relative policy optimizationgroup relative policy optimization, n.: A reinforcement-learning recipe for language models that has the model answer the same question many times and rewards each answer by how much better it did than the group's average. (GRPO), rewarding correct final answers to mathematics and programming problems and well-formatted output. Its precursor, R1-Zero, trained by reinforcement learning alone without supervised examples, developed long chains of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate., self-checking and backtracking on its own, showing that such behavior can emerge from outcome rewards. R1 matched OpenAI o1OpenAI o1, prop. n.: OpenAI's first reasoning model, from 2024. Before replying it writes out a long, hidden series of steps, trying ideas and checking them, which lifted its scores on hard math, science and coding problems far above earlier chatbots'. on several benchmarks, showed its full reasoning, and was released alongside smaller distilledknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run. versions built on QwenQwen, prop. n.: A family of AI language models from the Chinese company Alibaba, many free to download, which by 2025 had become the most popular starting point for other people's open AI models. and LlamaLlama, prop. n.: A family of AI language models from Meta, first released in 2023, whose trained settings can be downloaded and run by anyone, making them the foundation for thousands of other AI projects. base models. It set off the DeepSeek momentDeepSeek moment, n.: The shock of January 2025, when a little-known Chinese company released a free AI model nearly as good as the best American ones at a fraction of the reported cost, briefly wiping hundreds of billions off chip stocks. and made open reasoning models a standard target.
We distilled DeepSeek-R1's reasoning traces into a 7-billion-parameter model for on-device use.
Origin
Described in the technical report 'DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning' (January 2025) and in a peer-reviewed version in Nature (September 2025), which put the cost of R1's reasoning training at about US$294,000 on top of the base model. An updated R1-0528 followed in May 2025; DeepSeek's later V3.1, V3.2 and V4 (preview, April 2026) models folded reasoning into its main line.
Usage note
Like other Chinese models, its hosted versions avoid topics censored in China. Data-security concerns led several governments and agencies to bar its app from official devices.
Related
- DeepSeek momentDeepSeek moment, n.: The shock of January 2025, when a little-known Chinese company released a free AI model nearly as good as the best American ones at a fraction of the reported cost, briefly wiping hundreds of billions off chip stocks.
- reasoning modelreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning.
- reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds.
- group relative policy optimizationgroup relative policy optimization, n.: A reinforcement-learning recipe for language models that has the model answer the same question many times and rewards each answer by how much better it did than the group's average.
- OpenAI o1OpenAI o1, prop. n.: OpenAI's first reasoning model, from 2024. Before replying it writes out a long, hidden series of steps, trying ideas and checking them, which lifted its scores on hard math, science and coding problems far above earlier chatbots'.
- open-weight modelopen-weight model, n.: An AI model whose trained internal numbers are published for anyone to download and run on their own computers, even if the data and code used to make it stay private.
- DeepSeekDeepSeek, prop. n.: A Chinese artificial intelligence company whose free, cheaply trained models startled the world in January 2025 by nearly matching the best American systems.
Sources
- DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948; peer-reviewed version in Nature, September 2025.
Mentioned in
DeepSeekDeepSeek, prop. n.: A Chinese artificial intelligence company whose free, cheaply trained models startled the world in January 2025 by nearly matching the best American systems., knowledge distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run., LlamaLlama, prop. n.: A family of AI language models from Meta, first released in 2023, whose trained settings can be downloaded and run by anyone, making them the foundation for thousands of other AI projects., open-weight modelopen-weight model, n.: An AI model whose trained internal numbers are published for anyone to download and run on their own computers, even if the data and code used to make it stay private., reasoning modelreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning., DeepSeek momentDeepSeek moment, n.: The shock of January 2025, when a little-known Chinese company released a free AI model nearly as good as the best American ones at a fraction of the reported cost, briefly wiping hundreds of billions off chip stocks., gpt-ossgpt-oss, prop. n.: Two free-to-download reasoning models released by OpenAI in August 2025 — its first downloadable language models since 2019 — that anyone may run, modify and use commercially on their own hardware., group relative policy optimizationgroup relative policy optimization, n.: A reinforcement-learning recipe for language models that has the model answer the same question many times and rewards each answer by how much better it did than the group's average., Liang WenfengLiang Wenfeng, prop. n.: A Chinese entrepreneur who founded DeepSeek, whose low-cost, openly released AI models shocked global markets in January 2025., OpenAI o1OpenAI o1, prop. n.: OpenAI's first reasoning model, from 2024. Before replying it writes out a long, hidden series of steps, trying ideas and checking them, which lifted its scores on hard math, science and coding problems far above earlier chatbots'., QwenQwen, prop. n.: A family of AI language models from the Chinese company Alibaba, many free to download, which by 2025 had become the most popular starting point for other people's open AI models., reasoning tracereasoning trace, n.: The written-out 'thinking' an AI reasoning model produces before its final answer — notes to itself full of attempts, checks and corrections, which users may or may not get to see., reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds., RL scalingRL scaling, n.: The practice, and the finding behind it, of giving AI models far more trial-and-error practice after their initial training, especially on problems with checkable answers, which has so far improved their reasoning fairly predictably.