DeepSeek-R1
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Alignment and reasoning
- Lineage
- Symbols and logic
- Kind
- Implementation
Who
DeepSeek
What happened
The Chinese company DeepSeek releases DeepSeek-R1, a reasoning model with open weights under the MIT license. Its app overtakes ChatGPT as the most downloaded on the US App Store, and on January 27 Nvidia's shares fall 17 percent.
Why it mattered
It reproduces the reasoning-model paradigm with open weights. Its R1-Zero variant shows that reasoning can emerge from pure reinforcement learning.
The full story
DeepSeek, based in Hangzhou, was founded in July 2023 by Liang Wenfeng and funded by High-Flyer, a Chinese hedge fund. As early as 2021 Liang was buying Nvidia GPUs in bulk, eventually amassing 10,000 A100s before the United States restricted sales of those chips to China.
On January 20, 2025, the company released R1, built on its DeepSeek-V3 model, with the weights available under the MIT license. The paper presented two models. The first, R1-Zero, was trained with reinforcement learning alone, with no human-written reasoning examples: by rewarding only correct answers on verifiable math and coding problems, the model developed on its own behaviors such as reviewing its work, checking results, or switching strategies. But its output was hard to read and mixed languages. R1, the model released for use, started from a few thousand examples of long reasoning and then alternated reinforcement and supervised fine-tuning phases. The paper also showed that the reasoning patterns of a large model could be used to improve smaller ones.
The reaction was immediate. On January 27 its app overtook ChatGPT as the most downloaded free app on the US App Store, and Nvidia's shares fell 17 percent, a loss of $589 billion in market value, the largest one-day drop in US history. Part of the impact came from one figure: about $5.6 million, which DeepSeek had given as the cost of V3's final training run, not of the whole development.
The paper was later published in Nature. R1 showed that a Chinese lab, with an open model, could come very close to the best American reasoning models.
Under the hood
R1-Zero was trained with GRPO, a variant of PPO that needs no separate model to estimate value. For each problem, the model writes a group of answers; each one is scored by rules, for whether it is correct and for whether it respects the format that separates the reasoning from the answer, and its advantage is how far its score exceeds the group average.
With that alone, R1-Zero's accuracy on the 2024 AIME problems rose from 15.6 percent to 71.0 percent, and its answers grew longer and longer.
R1 went through four phases: fine-tuning on a few thousand long, readable reasoning traces; reinforcement like R1-Zero's, with an extra reward for not mixing languages; fine-tuning on about 800,000 samples selected from the model's own answers; and another round of reinforcement for general tasks. Those 800,000 samples were used to fine-tune Qwen and Llama models from 1.5 billion to 70 billion parameters.
Side by side
| Aspect | R1-Zero | R1 |
|---|---|---|
| Starting point | DeepSeek-V3 without fine-tuning | DeepSeek-V3 fine-tuned on thousands of written reasoning traces |
| Training | Reinforcement learning only | Alternating reinforcement and supervised fine-tuning |
| Output | Hard to read, with mixed languages | Readable |
| Purpose | Show that reinforcement is enough | Public use |
Common misconceptions
What people often say: DeepSeek trained a cutting-edge reasoning model for $5.6 million.
In fact: That is the figure DeepSeek gave for the final training run of V3, the base model, and it does not include the research, the earlier trial runs, or the reinforcement learning that turned V3 into R1.
Builds on
OpenAI o1 (September 12, 2024)
Inherited, with no proven direct link
The same paradigm; no direct dependence is documented
DeepSeek-V3 (December 26, 2024)
Documented direct influence
R1-Zero and R1 are trained from DeepSeek-V3-Base (section 2)
Leads to
DeepSeek V4 (April 24, 2026)
Inherited, with no proven direct link
V4 is the same lab's next major model, with the same MIT license and the same low-price strategy
Qwen3 (April 29, 2025), Kimi K2 (July 11, 2025)
Inherited, with no proven direct link
After R1, other Chinese labs release open frontier models under permissive licenses
Threads that pass through here
- From prompted reasoning to trained reasoning (step 6 of 9)
- China and the chips (step 3 of 6)
Sources
Original source
Scholarship and context
Further reading
How to cite
“DeepSeek-R1.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/deepseek-r1
This entry was last reviewed on November 13, 2026.