Chinese Ai Lab Deepseek published an open version of Deepseek-R1, his so-called reasoning model, which claims to play as well as that of Openai O1 On certain benchmarks of AI.
R1 is available from the AI Dev platform hugging the face under a MIT license, which means that it can be used commercially without restrictions. According to Deepseek, R1 Bat O1 on the AIME references, Math-500 and Swe-Bench verified. Like uses other models to assess the performance of a model, while Math-500 is a collection of word problems. Swe-Bench verified, meanwhile, focuses on programming tasks.
Being a reasoning model, R1 is effectively verified helps to avoid some of the traps that normally stumble the models. The reasoning models take a little more time – seconds to a few more minutes – to reach solutions compared to a typical non -rename model. The advantage is that they tend to be more reliable in fields such as physics, science and mathematics.
R1 contains 671 billion parameters, In depth revealed in a technical report. The parameters roughly correspond to the problem solving skills of a model, and the models with more parameters generally work better than those with fewer parameters.
Indeed, 671 billion parameters are massive, but Deepseek also published “distilled” versions of R1, from 1.5 billion parameters to 70 billion parameters. The smallest can work on a laptop. As for the complete R1, it requires more robust equipment, but it East Available via the Deepseek API at 90% to 95% cheaper than the O1 of Openai.
Clem Delangue, the CEO of Hugging Face, said in a Publish On Monday, the developers of the platform created more than 500 R1 “derivatives” models that accumulated 2.5 million combined downloads – five times the number of downloads that the official R1 has obtained.
There is a drawback at R1. Being a Chinese model, it is subject to reference By the Chinese Internet regulator to ensure that his answers “embody the basic socialist values”. R1 will not answer questions about Tiananmen Square, for example, or the autonomy of Taiwan.
A lot Chinese AI systemsincluding Other reasoning models,, decline To respond to subjects that could raise the anger of regulators in the country, such as speculation on Xi Jinping diet.
R1 arrives a few days after the outgoing Biden administration propose harder Export rules and restrictions on AI technologies for Chinese companies. Companies in China were already prevented from buying advanced AI chips, but if the new rules come into force, companies will face more strict ceilings on semiconductor technology and the models necessary for sophisticated bootstrap of the systems ‘Ia.
In a policy document Last week, Openai urged the United States government to support the development of the American AI, lest the Chinese models correspond to or exceed them in capacity. In a interview With the information, the vice-president of Openai’s policy, Chris Lehane, distinguished High Flyer Capital Management, the Deepseek business parent, as an organization of a particular concern.
Until now, at least three Chinese laboratories – Deepseek, Alibaba and Kimibelonging to Chinese unicorn Moshot ai – have produced models that they claim Rival O1. (It should be noted that Deepseek was the first – he announcement an overview of R1 at the end of November.) In a job On X, Dean Ball, AI researcher at George Mason University, said the trend suggests that Chinese AI laboratories will continue to be “fast followers”.
“The impressive performance of the DEEPSEEK distilled models (…) mean that very competent reasoners will continue to proliferate widely and run on local equipment,” wrote Ball, “far from the eyes of any downhill control diet.”
This story initially published on January 20 and was updated on January 27 with more information.
Techcrunch has a newsletter focused on AI! Register here To get it in your reception box every Wednesday.