← Back to the A-Z glossary
Model · DeepSeek · 2025

DeepSeek R1

DeepSeek R1 is the open-weight reasoning model DeepSeek shipped in January 2025. It was the first to deliver reasoning close to OpenAI's o1, but with weights published for free. The shock that showed open source could catch up to closed source way faster than anyone expected.

Model 4 min read Updated 2026-05-25
— What it is

DeepSeek R1, in plain words

In plain words: a model that takes the time to think before it answers (so it's stronger at math, code, and logic), but whose weights you can download for free, without paying OpenAI.

And then DeepSeek drops R1: performance very close to o1 on the benchmarks, weights published for free on Hugging Face, and a technical paper explaining how they pulled it off. That's the moment the world realized open source was catching up to closed source faster than expected.

R1 has been replaced in their app by newer versions (V3.x), but it's still downloadable and still used as a base for community spinoffs (R1-Distill, and so on).

What it's for

  • Local reasoning: running a reasoning model on my own machine without sending my data to OpenAI.
  • Math and logic: R1's reasoning chains hold up well for solving multi-step problems.
  • A base for fine-tuning: the community starts from R1 to build specialized models (legal, medical, and so on).
  • Learning how a model reasons: R1 shows its reasoning step by step, which makes it a great teaching tool.
  • A historical archive: to understand how open source caught up with OpenAI in 2025.

How it compares

vs DeepSeek V3.2: V3.2 is the current all-rounder, more versatile. R1 is still better at pure reasoning, but V3.2 has a "thinking" mode that gets close. Pick R1 if all you want is reasoning, V3.2 otherwise.

vs OpenAI o1 / o3: o1/o3 are closed, R1 is open. Performance was close in January 2025; OpenAI has pulled slightly ahead since, but R1 still holds up for plenty of uses.

vs Llama 4 Behemoth: Behemoth is stronger at reasoning, but far heavier (impossible to run locally on a consumer machine). R1 is more accessible.

What it costs

Prices as of May 25, 2026:

  • Model weights: free on huggingface.co/deepseek-ai
  • License: MIT (free, usable commercially)
  • Official DeepSeek API: ~$0.55 / million tokens (input) — pricier than V3.2 because reasoning burns more
  • Via Fireworks or Together: $1-3 / million tokens depending on the provider

My take

I don't use R1 directly anymore — V3.2 is better for most of my cases and costs less. But R1 is still the founding moment of 2025, the one that proved open source can produce cutting-edge models.

If you're just getting into AI in 2026, you don't need to use R1. If you do research, or you want to understand how open-source AI exploded, it's a reference worth knowing.

Its real value today: the base that dozens of community models were built on.

Quick questions

Is DeepSeek R1 outdated?

DeepSeek R1 is the open-weight reasoning model released by DeepSeek in January 2025, under an MIT license. The first o1-style reasoning with free weights, it marked the moment open source caught up.

Can I run it on my MacBook?

Not the full R1 (671 billion parameters). But distilled versions (7B, 32B) run on a MacBook M3/M4.

Is R1 free for commercial use?

Yes, MIT license, usable with no restrictions.

Checked on 2026-05-25 · next review 2026-11-25

Specs and prices checked on Hugging Face and the official DeepSeek docs. Page documented as an archive and to help you understand the open-weight ecosystem.

Want to keep going in order?

Head back to the full glossary, search a word, then open only the pages that deserve more than a short definition.

Open the AI glossary