DeepSeek R1 Explained: How It Was Trained, Benchmarks vs OpenAI o1, Distilled Models and Costs
What DeepSeek R1 is and why it mattered: the R1-Zero reinforcement learning experiment, the multi-stage training behind R1, published benchmarks against OpenAI o1, the distilled 1.5B to 70B models, MIT licensing, API pricing at launch and how to run it.
DeepSeek R1, released in January 2025 by the Chinese lab DeepSeek, was the first openly licensed reasoning model to report results comparable to OpenAI’s o1 on maths and coding benchmarks. It mattered for three reasons: its training paper showed that reinforcement learning alone could produce step-by-step reasoning, the weights were released under the permissive MIT license, and its API launched at a small fraction of o1’s price. This guide explains how R1 was built, what its published numbers show, which version to use and what its limits are.
Key facts
| DeepSeek R1 | |
|---|---|
| Released | January 2025 |
| Architecture | Mixture-of-experts, 671B total parameters, 37B active per token |
| Context length | 128K tokens |
| License | MIT (commercial use, modification and distillation allowed) |
| Distilled versions | 1.5B, 7B, 8B, 14B, 32B and 70B, based on Qwen2.5 and Llama 3 |
| Paper | DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning |
Specifications are from the DeepSeek-R1 model card on Hugging Face.
How R1 was trained
DeepSeek actually published two models. The first, DeepSeek-R1-Zero, was trained with large-scale reinforcement learning directly on the DeepSeek-V3 base model, with no supervised fine-tuning step first. Rewarded for reaching correct, checkable answers in maths and code, it learned on its own to produce long chains of reasoning, check its work and revisit mistakes. That was the scientifically notable result: reasoning behaviour emerged without human-written demonstrations of how to reason.
R1-Zero had practical problems, however. Its output could be repetitive, hard to read and mix languages. To fix this, DeepSeek-R1 added a small amount of curated “cold-start” data before reinforcement learning and then went through further training stages. So the commonly repeated claim that R1 itself used “pure reinforcement learning with no supervised fine-tuning” describes R1-Zero, not the model most people use.
What the benchmarks show
DeepSeek’s model card compares R1 with OpenAI’s o1 (December 2024 version) and other models of the time:
| Benchmark | DeepSeek R1 | OpenAI o1-1217 | OpenAI o1-mini |
|---|---|---|---|
| AIME 2024 (Pass@1) | 79.8 | 79.2 | 63.6 |
| MATH-500 (Pass@1) | 97.3 | 96.4 | 90.0 |
| Codeforces rating | 2029 | 2061 | 1820 |
| LiveCodeBench (Pass@1, CoT) | 65.9 | 63.4 | 53.8 |
| SWE-bench Verified (resolved) | 49.2 | 48.9 | 41.6 |
These are DeepSeek’s own reported figures. They show R1 roughly matching o1 on competition maths and coding rather than clearly beating it. The paper also acknowledges weaknesses: R1 was less capable than DeepSeek-V3 at some general tasks such as function calling, multi-turn conversation and structured output, and it was sensitive to prompting, with few-shot prompts tending to hurt performance.
Which version to use
The full 671B model needs a multi-GPU server; it is not something most individuals can run. The distilled models are smaller open models (Qwen2.5 and Llama 3 bases) fine-tuned on reasoning examples generated by R1. They inherit much of its step-by-step style at far lower hardware cost.
| Option | Best for | Hardware |
|---|---|---|
| Hosted API (DeepSeek or third-party providers) | Trying R1 or production use without infrastructure | None |
| Distill 7B / 8B / 14B | Local experiments, laptops and single consumer GPUs | Consumer GPU or Apple silicon, especially with quantization |
| Distill 32B / 70B | Stronger local reasoning | High-memory GPU or workstation |
| Full R1 | Self-hosted production at scale | GPU cluster |
Tools such as Ollama, LM Studio and Hugging Face Transformers can run the distilled versions; community quantizations trade some quality for lower memory use. A sensible approach is to start with a hosted API or a small distilled model and move up only when your tasks need it.
Costs and licensing
At launch, DeepSeek listed R1’s API at $0.55 per million input tokens (cache miss) and $2.19 per million output tokens, against OpenAI’s $15 and $60 for o1 at the time. Prices have changed since; check DeepSeek’s current pricing page before planning a budget.
The MIT license allows commercial use, modification and building proprietary products on top, including distilling R1 into other models. Self-hosting removes per-token fees but replaces them with hardware, energy and engineering costs, which only pay off with steady, high-volume workloads.
Practical uses and limits
R1’s visible reasoning makes it useful where you want to inspect how an answer was reached: maths and science problems, debugging and explaining code, and analytical tasks where checking intermediate steps helps. Because the reasoning is long, responses are slower and use more output tokens than a standard chat model, so it is a poor fit for quick, simple requests.
Two cautions apply. First, the visible reasoning can still contain errors, so treat it as something to check rather than proof. Second, the open weights and DeepSeek’s hosted app are different things: running the model yourself keeps data on your infrastructure, while the hosted service is subject to DeepSeek’s own terms and data handling, which some organisations and governments have restricted.
What came next
R1 pushed other labs to cut prices and release more open reasoning models, and DeepSeek itself has since shipped newer versions. R1 remains a useful reference point for understanding how reinforcement learning on verifiable rewards became a standard part of training reasoning models.
Common questions
Is DeepSeek R1 really open source? The weights are released under the MIT license and the training approach is described in the paper, but the training data and full training code are not published. “Open-weight” is the more precise term.
How should I prompt R1? DeepSeek’s paper reports that few-shot examples tend to degrade its performance. It recommends describing the problem directly and specifying the output format in a zero-shot prompt.
Does the visible reasoning mean the answer is correct? No. The reasoning makes errors easier to spot, but it can contain mistakes or reach a wrong conclusion confidently, so verify important results independently.
Why do some organisations block it? Restrictions have usually targeted DeepSeek’s hosted app and its data handling rather than the open weights, which can run entirely on infrastructure you control.
Sources and further reading
Our analysis
R1’s real significance was methodological and economic rather than a decisive benchmark win. It showed that reinforcement learning on verifiable rewards can produce strong reasoning, published the recipe openly, and priced access low enough to change market expectations. For most teams, the distilled models and hosted APIs are the practical entry points, and data-handling questions about the hosted app should be considered separately from the open weights.