DeepSeek drops R1, open-sourcing an o1-level reasoning model under MIT

DeepSeek-R1 is a fully open-source reasoning model built to tackle complex math, multi-step logic, and software engineering at the level of OpenAI-o1.

What it does and how it works

DeepSeek-R1 specializes in deep reasoning tasks. In benchmark evaluations across advanced mathematics, coding, and logical deduction, it delivers results directly comparable to OpenAI-o1.

Under the hood, the system relies on large-scale reinforcement learning applied during post-training. By heavily leaning on RL rather than massive human-annotated datasets, the model learned to explore solutions and refine its own reasoning chains with minimal supervised supervision.

Alongside the flagship weights, the team released six distilled smaller models. The 32B and 70B variants punch well above their weight, matching OpenAI-o1-mini on competitive math and programming benchmarks while remaining far easier to host.

Licensing, models, and pricing

DeepSeek made the entire suite available under the ultra-permissive MIT license. This covers model weights, technical reports, and the distilled model family.

Crucially, the release terms explicitly permit using DeepSeek-R1 weights and API outputs for downstream fine-tuning and model distillation. This removes the restrictive data clauses typically found in proprietary frontier APIs.

Anyone can test the model immediately on chat.deepseek.com by toggling the DeepThink feature. Developers can access it through the official API under the deepseek-reasoner endpoint.

API pricing is aggressive. Input tokens cost $0.14 per million on cache hits and $0.55 per million on cache misses, while output tokens are priced at $2.19 per million.

Why it matters

If you build software, run automated workflows, or need heavy logic reasoning, DeepSeek-R1 is an immediate must-try. You no longer have to pay premium rates for proprietary frontier endpoints just to get chain-of-thought problem-solving, and you can self-host the 32B or 70B versions locally without asking for permission.

By releasing full weights under MIT and greenlighting output distillation, DeepSeek effectively gave every developer the building blocks to train specialized reasoning engines. It turns frontier reasoning from a proprietary service into a widespread open utility.