DeepSeek Releases R1: o1-Style Reasoning, Open and MIT-Licensed

Chinese lab DeepSeek has released R1, a reasoning model that, in the company's own words, performs "on par with OpenAI-o1" on math, coding and reasoning tasks. The difference from the American benchmark isn't in raw capability, but in how it's distributed: R1 ships with the model and full technical report completely open, under an MIT license, and with API pricing well below what's typical for models in this category.

What a reasoning model is, and why it matters

The family of "reasoning" models — the one OpenAI kicked off with o1 — differs from a conventional chatbot in that it spends extra compute time "thinking" before answering, chaining together intermediate steps instead of firing off the first response that comes to mind. That approach particularly boosts performance on math, coding and logic problems, where a single wrong step can wreck the final answer.

Until now, that kind of capability had been locked behind closed, paid models accessible only through their maker's API. R1 breaks that exclusivity: DeepSeek is releasing it with open weights — the internal parameters that define the model — under one of the most permissive licenses available.

The MIT license changes everything

The legal fine print is what carries the biggest consequences here. DeepSeek has put R1 under an MIT license, which in practice means anyone can use it, modify it and commercialize it freely. The company put it bluntly: "Distill & commercialize freely!"

That last point is no small detail. DeepSeek specifies that API outputs can now be used for fine-tuning & distillation — that is, to fine-tune other models or to train smaller versions from R1's responses. Many closed-model providers explicitly ban using their outputs to train competing systems; here, it's explicitly allowed.

For the open-source community, this turns R1 into a foundation to build on without asking permission or paying for licenses — something that, until recently, seemed reserved for large labs with billion-dollar budgets.

Six distilled models for more modest hardware

Alongside the main model, DeepSeek has released six smaller distilled models, all open as well. Distillation transfers the behavior of a large model into a smaller one, so the latter retains much of the capability while requiring far fewer resources.

According to the company, its 32-billion and 70-billion-parameter versions perform "on par with OpenAI-o1-mini," the lightweight variant of OpenAI's reasoning model. These distilled versions dramatically widen the potential audience: not everyone can run a full frontier model, but a mid-sized one opens the door for companies, universities and independent developers to work with advanced reasoning on more affordable infrastructure.

How it was trained: large-scale reinforcement learning

The method is documented in a technical report DeepSeek published alongside the model. The key, according to the company, lies in using large-scale RL in post-training — the fine-tuning phase that follows base training.

Reinforcement learning is a technique in which the model improves through trial and reward, adjusting its behavior based on the quality of its responses. DeepSeek states that this approach achieved a "significant performance boost with minimal labeled data." That detail matters: hand-labeling data is expensive and slow, and cutting that dependency speeds up and cheapens development. The fact that the lab is publicly sharing its recipe also lets others reproduce and scrutinize it — something rare among leading closed models.

An API at a fraction of the usual price

The other blow lands on cost. R1 is already available via API under the model name deepseek-reasoner, priced as follows:

• $0.14 per million input tokens on a cache hit.
• $0.55 per million input tokens on a cache miss.
• $2.19 per million output tokens.

Tokens are the units models use to process text, and caching lowers the cost of requests that reuse content already processed. For a reasoning model — which burns through plenty of tokens precisely because it "thinks" before answering — this pricing level lowers the economic barrier to testing it at scale. The model can also be tried through the company's web interface, using its DeepThink reasoning mode.

What this release calls into question

The dominant narrative of recent years has held that reaching the frontier of AI required colossal budgets, hard-to-access hardware and teams only a handful of companies could afford. R1 challenges that premise on three fronts at once: it matches — according to DeepSeek — a benchmark model like o1, it does so in the open with a free commercial license, and it's priced well below equivalent services.

The performance comparisons should be taken with some caution, since for now they come from the maker itself and will need to be confirmed by independent evaluations. But even with that caveat, the move has clear implications. For developers and companies, there's now an alternative they can inspect, modify and deploy without restrictions. For labs defending the closed-model approach, the gap between proprietary and open narrows. And for the open-source ecosystem, R1 and its distilled versions bring a level of advanced reasoning that, until recently, didn't exist outside paid platforms.

The fact that a Chinese lab is publishing its full method and releasing the weights adds a geopolitical dimension to the debate over who controls frontier AI and under what terms it gets distributed. How the major providers respond — on price, on openness, or both — will shape the months ahead.