Skip to content
WAR OF THE WEIGHTS

Cutting-edge Chinese “reasoning” model rivals OpenAI o1—and it’s free to download

DeepSeek R1 is free to run locally and modify, and it matches OpenAI's o1 in several benchmarks.

Benj Edwards | 209

On Monday, Chinese AI lab DeepSeek released its new R1 model family under an open MIT license, with its largest version containing 671 billion parameters. The company claims the model performs at levels comparable to OpenAI's o1 simulated reasoning (SR) model on several math and coding benchmarks.

Alongside the release of the main DeepSeek-R1-Zero and DeepSeek-R1 models, DeepSeek published six smaller "DeepSeek-R1-Distill" versions ranging from 1.5 billion to 70 billion parameters. These distilled models are based on existing open source architectures like Qwen and Llama, trained using data generated from the full R1 model. The smallest version can run on a laptop, while the full model requires far more substantial computing resources.

The releases immediately caught the attention of the AI community because most existing open-weights models—which can often be run and fine-tuned on local hardware—have lagged behind proprietary models like OpenAI's o1 in so-called reasoning benchmarks. Having these capabilities available in an MIT-licensed model that anyone can study, modify, or use commercially potentially marks a shift in what's possible with publicly available AI models.

"They are SO much fun to run, watching them think is hilarious," independent AI researcher Simon Willison told Ars in a text message. Willison tested one of the smaller models and described his experience in a post on his blog: "Each response starts with a <think>...</think> pseudo-XML tag containing the chain of thought used to help generate the response," noting that even for simple prompts, the model produces extensive internal reasoning before output.

Ars Video

 

Simulated reasoning in action

The R1 model works differently from typical large language models (LLMs) by incorporating what people in the industry call an inference-time reasoning approach. They attempt to simulate a human-like chain of thought as the model works through a solution to the query. This class of what one might call "simulated reasoning" models, or SR models for short, emerged when OpenAI debuted its o1 model family in September 2024. OpenAI teased a major upgrade called "o3" in December.

Unlike conventional LLMs, these SR models take extra time to produce responses, and this extra time often increases performance on tasks involving math, physics, and science. And this latest open model is turning heads for apparently quickly catching up to OpenAI.

For example, DeepSeek reports that R1 outperformed OpenAI's o1 on several benchmarks and tests, including AIME (a mathematical reasoning test), MATH-500 (a collection of word problems), and SWE-bench Verified (a programming assessment tool). As we usually mention, AI benchmarks need to be taken with a grain of salt, and these results have yet to be independently verified.

A chart of DeepSeek R1 benchmark results, created by DeepSeek.
A chart of DeepSeek R1 benchmark results, created by DeepSeek. Credit: DeepSeek

TechCrunch reports that three Chinese labs—DeepSeek, Alibaba, and Moonshot AI's Kimi—have now released models they say match o1's capabilities, with DeepSeek first previewing R1 in November.

But the new DeepSeek model comes with a catch if run in the cloud-hosted version—being Chinese in origin, R1 will not generate responses about certain topics like Tiananmen Square or Taiwan's autonomy, as it must "embody core socialist values," according to Chinese Internet regulations. This filtering comes from an additional moderation layer that isn't an issue if the model is run locally outside of China.

Even with the potential censorship, Dean Ball, an AI researcher at George Mason University, wrote on X, "The impressive performance of DeepSeek's distilled models (smaller versions of r1) means that very capable reasoners will continue to proliferate widely and be runnable on local hardware, far from the eyes of any top-down control regime."

Photo of Benj Edwards
Benj Edwards Senior AI Reporter
Benj Edwards is Ars Technica's Senior AI Reporter and founder of the site's dedicated AI beat in 2022. He's also a tech historian with almost two decades of experience. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.
209 Comments
Staff Picks
Artem S. Tashkinov
I’m sure stealing technology from other companies helps accelerate their progress.
One thing is stealing.

The other thing is making it work and making it work even better than the original.

That's why the whole AI community has been buzzing about DeepSeek. Its training was estimated to cost a tiny fraction of what e.g. OpenAI has spent and not only that it's a lot cheaper to run.

You cannot underestimate the potential of 1.5 billion people.
AI is cool i guess
Oh. OH.

this isn't just another open source LLM release. this is o1-level reasoning capabilities that you can run locally. that you can modify. that you can study. that's...
that's a very different world than the one we were in yesterday.

(and the fact that it's coming from china and it's MIT licensed? the geopolitical implications here are fascinating)

but the really wild part? those distilled models. we're talking about running reasoning models on consumer hardware. remember when everyone said this would be locked up in proprietary data centers forever?

something absolutely fundamental just shifted in the AI landscape. again, this is getting intense.

(also, wouldn't it be wild if deepseek renamed themselves to ClosedAI? 🤣)

2025 is going to be wiiiild