The Chinese Deepseek laboratory has released an open version of Deepseek-R1, its so-called reasoning model, which states that it performs and O1 of Openi on some reference parameters of the AI.
R1 is available from the platform to the Devs that embraces the face with a MIT license, which means that it can be used commercially without restrictions. According to Deepseek, R1 beats O1 on Aime’s benchmark, Math-500 and Swe-Bench verified. Aime uses other models to evaluate the performance of a model, while Math-500 is a collection of words of words. In the meantime, the SWE bench is verified on programming tasks.
Being a reasoning model, R1 occurs effectively itself, which helps it to avoid some of the pitfalls that normally stumble on the models. The reasoning models require a little more time, usually seconds or minutes longer – to get to solutions than a typical non -seasonal model. The positive side is that they tend to be more reliable in sectors such as physics, science and mathematics.
R1 contains 671 billion of parameters, Deepseek revealed in a technical relationship. The parameters match approximately to the resolution skills of the problems of a model and models with multiple parameters generally work better than those with less parameters.
In fact, 671 billion parameters are enormous, but Deepseek has also released “distilled” versions of R1 of size between 1.5 billion parameters to 70 billion parameters. The smallest can work on a laptop. As for the entire R1, it requires more robust hardware, but it AND Available through the Deepseek API at 90% -95% cheaper prices than OPEN.
Clem Delague, CEO of Hugging Face, declared in an X post on Monday that the developers on the platform created more than 500 “derivatives” derivative models that collected 2.5 million combined downloads, five times the download number of official R1 is obtained.
There is a negative aspect for R1. Being a Chinese model, it is subject to benchmarking by the Chinese internet regulator to ensure that its responses “embodies the basic socialist values”. R1 will not answer questions about Tiananmen Square, for example, or Taiwan’s autonomy.
Many Chinese artificial intelligence systems, including other reasoning models, refuse to respond to the topics that could increase the anger of regulators in the country, such as speculation on the XI Jinping regime.
R1 arrives days after the outgoing administration Biden has proposed rules and restrictions of more severe export to artificial intelligence technologies for Chinese initiatives. Companies in China had already been prevented from buying chips at the advanced, but if the new rules enter into force as writings, companies will have to face more rigorous hats both on the technology of semiconductors and on the models necessary for bootstrap sophisticated artificial intelligence systems.
In a political document of last week, Openai urged the United States government to support the development of AI in the United States, so as not to pay the Chinese models or overcome them in capacity. In an interview with information, the vice -president of Open by Openi Chris Lehane has identified High Flyer Capital Management, the company’s parent of Deepsiek, as an organization of particular concern.
So far, at least three Chinese workshops – Deepseek, Alibaba and Kimi, owned by a Chinese Moon Shot Unicorn, have produced models that support the rival O1. (Note, Deepseek was the first: he announced a preview of R1 in late November.) In a post on X, Dean Ball, an artificial intelligence researcher at the George Mason University, said that the trend suggests that i Ai Chinese workshops will continue to be “fast followers.”
“The impressive performance of the distilled models of Deepseek (…) means that very capable accountants will continue to proliferate widely and be executable on the local hardware”, wrote Ball, “away from the eyes of any control regime from above down. “
This story was originally published on January 20 and was updated on January 27 with further information.
Techcrunch has a newsletter focused on artificial intelligence! Sign up here To do it in your mailbox every Wednesday.
