DeepSeek-R1 arrived in January 2025 from a Chinese lab and climbed to the top of app store charts within days. It shows its reasoning before answering and was released with open weights.
AI model data sheet: DeepSeek-R1
| Developer | DeepSeek |
|---|---|
| Released | January 20, 2025 |
| Type | Reasoning model (mixture-of-experts, 671B total parameters) |
| Context window | 128K tokens |
| Access | DeepSeek app, API and downloadable weights |
| Licence / weights | Open weights, MIT licence |
| Best at | Step-by-step maths, code and logic problems |
| Why it matters | Caused a global stir by matching top reasoning models at far lower reported training cost |
Quick verdict: A turning point for open reasoning models. Review data-privacy terms before using the hosted app.
Its release forced a rethink of how much money cutting-edge AI needs and sparked debate in markets and policy circles.