Availability · January 20, 2025

DeepSeek-R1

A Chinese laboratory opened a model that reasons at the level of o1 and showed that the behaviour emerges from pure reinforcement learning on the correct answer.

Why it matters

Reasoning stopped being one company's closed technology: the recipe turned out to be short and the weights were in the open.

What mattered was not the result but the method. The base model was trained by reinforcement on problems with a checkable answer, with no human example of reasoning at all, and it developed long chains with self-checking and backtracking by itself. The report described the moment during training when the model stops and rewrites its approach. Unlike o1, the chain was shown in full. The weights were published under an MIT licence, and dozens of smaller models grew on them within weeks.

Event record

Event date

January 20, 2025

Timeline date

Event date

Verification

Sources gathered automatically · September 18, 2026

Lines

ID

evt-0318

The weights and technical report were released on 20 January 2025.

Sources

primary DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
arXiv · Published January 2025
primary DeepSeek-R1 Release
DeepSeek · Published January 2025
secondary DeepSeek
Wikipedia

Related events

Extends DeepSeek-V3
The reasoning model is built on the same base released three weeks earlier.
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Related o1 and compute at answering time
The same result, but with open weights and a visible chain of reasoning.

Records that link to this one

Related Seventeen per cent in a day
The market was reacting to the model released a week earlier.
Related Qwen3
The second Chinese laboratory in a quarter to put frontier weights in the open.
Related gpt-oss
An answer to a year in which open weights became a strong argument in someone else's hands.
Related The United States opens H200 sales to China
The restrictions were imposed against laboratories that reached the frontier regardless.
Extends DeepSeek-V4
The fourth generation of the same line, fifteen months after the reasoning model.
DeepSeek-V4 Preview Release
Enables GRPO: reinforcement learning without a critic
GRPO, described here, is the algorithm with which DeepSeek-R1 was trained a year later.
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Contrasts A thousand examples and a forced budget
The paper's table sets s1-32B beside the R1 distillations, trained on 800 thousand examples against a thousand here.
s1: Simple test-time scaling (arXiv:2501.19393v1)
Related Long context inside reinforcement learning
Two days after DeepSeek-R1 appeared, another laboratory described its own route to the same thing: long context inside the reinforcement learning itself.
Related Search inside the chain of reasoning
The work is built around what long-reasoning models lack: during the reasoning itself they run into gaps in their knowledge.
Builds on Five thousand puzzles instead of a corpus
The paper's abstract opens with "Inspired by the success of DeepSeek-R1"; the work reproduces its recipe on a controlled corpus.
Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning (arXiv:2502.14768v1)