DeepSeek-R1 Reasoning Model
Summary: On January 20, 2025, DeepSeek launched DeepSeek-R1, matching Western reasoning benchmarks on math and coding using MLA and GRPO algorithms at a fraction of the traditional training cost.
In early 2025, the technology world experienced what analysts called the "DeepSeek Shock." On January 20, 2025, in Hangzhou, China, DeepSeek launched a reasoning model named DeepSeek-R1. Like OpenAI's o1, R1 was designed to "think" before answering, writing out a hidden list of logical steps. However, while Western models required billions of dollars in training resources, DeepSeek trained R1 for just $5.58 million. By open-sourcing the model and sharing its training recipes, DeepSeek proved that advanced logical reasoning could be built cheaply, disrupting the economics of global AI development.
| Historical Attribute | Milestone Registry Value |
|---|---|
| Classification Type | software |
| Chronological Date | 2025-01-20 |
| Coordinates / Location | Hangzhou, China |
| Curation Authority | Nick Hodder + MIA |
| Milestone Importance | godfather Milestone |
How does DeepSeek-R1 Reasoning Model fit into the history of artificial intelligence?
The transition to reasoning models was pioneered by the OpenAI o1 Reasoning Model Series in 2024, which introduced reinforcement learning to generate hidden chains of thought.
DeepSeek-R1 advanced this path by proving that Hebbian-like logical reinforcement could be achieved without massive proprietary datasets. Its release open-sourced these "thinking patterns," allowing researchers around the world to study and replicate machine reasoning locally, modifying the landscape of academic computer science.
What are the core technical achievements of DeepSeek-R1 Reasoning Model?
DeepSeek-R1 introduced "Group Relative Policy Optimization" (GRPO), a reinforcement learning algorithm that eliminates the need for a separate, memory-intensive critic model during training. This mathematically reduced the computing memory required, allowing the model to be trained on approximately 2,000 Nvidia H800 GPUs over 55 days.
The model's reasoning process is trained using reinforcement learning to reward correct answers in math and coding tasks. R1 automatically generates multi-step reasoning steps, showing comparable scores to closed proprietary systems on advanced benchmarks like AIME and Codeforces.
Why is the legacy of DeepSeek-R1 Reasoning Model significant to modern computing?
The legacy of DeepSeek-R1 lies in its economic disruption of the tech sector. By proving that reasoning capabilities do not require multi-billion-dollar computing clusters, it drove global AI labs to prioritize algorithmic efficiency and open-weight architectures.
Additionally, R1's release accelerated the integration of reasoning models in developer workflows, enabling local code generation, complex document analysis, and autonomous research tools to run cost-effectively, redefining the boundaries of computing.