DeepSeek

Contribution to AI

Reasoning models arrived in 2024 as a closed capability: OpenAI's o1 showed that spending more compute at inference time improved answers, but hid the chain of thought and said little about how the model had been trained to produce it. R1, released in January 2025 under an MIT licence with its method described in full, closed that gap in a fortnight. The sharper result was R1-Zero, which was given no worked examples at all — only a base model, a reward for correct answers, and a reward for putting its working between tags. Backtracking, self-checking and longer deliberation on harder problems appeared on their own, which settled an open question about whether reasoning traces had to be demonstrated by humans first. The V3 base model underneath made the second argument. A 671-billion-parameter mixture of experts with 37 billion active per token, trained in 2.788 million H800 GPU hours on export-restricted hardware, it used FP8 precision and a custom pipeline schedule to keep communication from dominating — engineering aimed at constrained chips rather than more of them. The market read this as a threat to the assumption that capability tracks capital, and Nvidia lost 593 billion dollars of value in a single session. The R1 paper later went through peer review at Nature, unusually for a frontier model.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).