Overview

DeepSeek-R1, released in January 2025 by the Chinese lab DeepSeek, showed that reinforcement learning on tasks with verifiable answers can teach a model long chains of reasoning. Its developers reported performance comparable to OpenAI’s o1 on math, coding and reasoning benchmarks. The open weights, MIT license and low API prices made it one of the most discussed AI releases of 2025.

Highlights

  • Mixture-of-experts architecture: 671B total, about 37B active per token
  • Visible step-by-step reasoning before answers
  • Trained with large-scale reinforcement learning
  • MIT license, including use for distillation
  • Distilled smaller versions based on Qwen and Llama

Typical uses

Math, coding, scientific reasoning, and research on reasoning methods; distilled versions for local use.

Things to consider

  • The full model requires large multi-GPU systems.
  • Long reasoning traces increase latency and token usage.
  • Some organizations restrict use of DeepSeek’s hosted services for data-governance reasons; self-hosting avoids sending data to the provider.

Specifications are taken from the provider’s official announcement, documentation or model card. Benchmark results depend on evaluation settings; see how AI benchmarks work.