What is DeepSeek-R1?

Yesterday DeepSeek, an AI lab based in China, released DeepSeek-R1, a model that thinks before it answers in the same spirit as OpenAI's o1 series. DeepSeek says R1 performs on par with o1 across math, coding and reasoning tasks. What makes the release stand out is not only the claimed performance but the terms: the model weights and a technical report are public, under the permissive MIT license, and DeepSeek explicitly allows people to use the model's outputs to train other models.

The model is already available in three ways. You can chat with it for free at chat.deepseek.com by switching on the "DeepThink" option. Developers can call it through DeepSeek's API under the name deepseek-reasoner. And anyone with the hardware can download the weights from Hugging Face.

How DeepSeek trained R1

DeepSeek actually describes two models. The first, DeepSeek-R1-Zero, was trained with reinforcement learning alone, without the usual first step of supervised fine-tuning on examples written by people. Reinforcement learning means the model tries many answers and is rewarded for good outcomes. According to DeepSeek, R1-Zero developed useful habits on its own, such as checking its work and reflecting on its steps, and the lab calls it the first open research to show reasoning can be encouraged purely this way.

R1-Zero had problems, though: it repeated itself endlessly, was hard to read, and mixed languages. So DeepSeek built R1 by adding a small amount of curated "cold-start" data before the reinforcement learning stages, plus further fine-tuning rounds. Both models are built on DeepSeek-V3-Base, the large open model the lab released on Christmas Day.

How big it is, and what it costs

R1 is a mixture-of-experts model: of its 671 billion parameters, only about 37 billion are active for any given token, which keeps it cheaper to run than its size suggests. It handles a context of up to 128,000 tokens.

The API prices are low by current standards:

A token is a small chunk of text, often part of a word. Because reasoning models write long internal chains of thought, output tokens are where most of the cost goes.

The distilled family

Alongside the big model, DeepSeek released six smaller ones. Each is an existing open model, from Alibaba's Qwen2.5 or Meta's Llama families, fine-tuned on reasoning examples generated by R1. This process is called distillation: a large "teacher" model's behavior is passed to a smaller "student." The sizes are 1.5, 7, 8, 14, 32 and 70 billion parameters. DeepSeek says the 32B and 70B versions are on par with OpenAI's o1-mini on its benchmarks, and that the 32B Qwen-based model sets a new high for models of its kind.

One licensing detail deserves attention: each distilled model also inherits the license of the model it started from, Apache 2.0 for the Qwen versions and Meta's Llama licenses for the Llama versions. Check those terms before building a product on one.

DeepSeek also publishes a few usage tips that differ from what many people do with chat models. It recommends a temperature (a randomness setting) between 0.5 and 0.7, with 0.6 as the default, to avoid endless repetition. It advises against a system prompt: put all instructions in your own message. For math, it suggests asking for the final answer in a box:

A train leaves at 9:40 and arrives at 13:05. A second train covers the same route 25 minutes faster. How long does the second train take? Put your final answer within \boxed{}.

Can you run an open reasoning model on your own computer?

Not the full R1, for most people. The complete model weighs in at more than 650 GB, far beyond any laptop. Running it takes server-class hardware, and DeepSeek points people to its V3 repository for instructions.

The distilled models change the picture. Developer Simon Willison ran an 8B distilled version on his Mac on release day, using a compressed ("quantized") copy prepared by the company Unsloth and the free tool Ollama. That download was about 8.5 GB. A day later he ran a quantized 70B version, a 34 GB download, on a Mac with 64 GB of memory. The small models print their reasoning between <think> tags before the answer, so you can watch them work.

A practical rule of thumb: start with the smallest model your machine handles comfortably, and move up only if the answers are too weak for your task. Expect the small versions to be noticeably less capable than the full R1 you can use online.

Running locally has real benefits. Your prompts never leave your computer, it works offline, and there is no per-token bill. The trade-offs are speed, quality and setup time. For many people the best path is to try R1 through the website first, then experiment with a distilled model at home if privacy or cost matters to them.