Hugging Face researchers are trying to build a more open version of DeepSeek’s AI ‘reasoning’ model

DEPSEC one week after it was revealed R 1 “logic” AI model – which Sending the market to a TG – Hugged face researchers are trying to replicate the model from scratch in what he says to follow “Open Knowledge”.

Liendro von Warrara and several companies’ engineers are involved in the research face Open-R 1 has launchedA project that it wants to create all its ingredients, including data used for training, to create a duplicate of R1 and open source.

Engineers said they were forced to work by the vision of the DIPC’s “Black Box”. Technically, the R1 “exposed” model is allowed to be allowed to be allowed to be approved, which means it can be originally placed without restrictions. However, no more 1 “Open source“According to the widely recognized definition because some of the tools used to make it are in the mystery. Like many high -flying AI companies, DIPSEC hates to publish the secret sauce.

“The R1 model is impressive, but Eli Buckchch, one of the Open-R 1 project hugs, told Techchen,” The R1 model is impressive, but no open datasate, test details or intermediate models are available, which makes the transcript and more research difficult. “” The complete architecture of complete open sourcing and 1 is not just about transparency – this is about unlocking its potential. “

So not open

A Chinese AI Lab DIPSEC, which is funded in the quantative hedge fund, released the R1 last week. At a number of criteria, R 1 matching – and even out of – Openai’s performance O 1 The logic of the logic.

Being a reasonable model, R1 effectively performs itself true-test, which It helps to avoid some problems that usually travel in modelsThe Reasonable models take a little longer to reach the solutions than a simple non-resoning model-for a few minutes more than a few minutes. The opposite is that they become more reliable in domains such as physics, science and mathematics.

After the Chattbot app of DIPSC, R1 broke the mainstream consciousness, which provides free access to R11, The Apple App Store has risen to the top of the chartThe R1 that was developed with speed and efficiency – a few weeks after the publication of the OpenAEA and 1 published the DEPSEC model – – Wall Street has led analysts And technicians Asked if the United States can maintain AI leadership.

The Open-R 1 project “Model Training Black Box is less concerned about our AI domination than fully opening,” Bakuch told TechCrunch. He mentioned that since the R11 training code or training directions were not published, the model of the model to study depthly – the behavior of the model is rarely driven.

“It is important to control the datasate and process for deploying a model in sensitive cases,” Buckch said. “It helps to understand and address the bias in the model. Researchers need more than pieces to stop the boundaries of what is possible. “

Transcript step

The goal of the Open-R 1 project is to make the R1’s transcript within a few weeks, hugged face science cluster, depending on a dedicated research server with 768 Nvidia H1 GPU.

The hug engineers plan to tap the science cluster to create a similar datasate used to create R1. To create a training pipeline, the team is seeking the help of AI and broad technology community in embracing face and githab, where the Open-R1 project is being hosted.

“We need to make sure we apply algorithm and recipes [correctly,]”Von Warra tells TechCrunch,” But it is perfect to deal with a community’s efforts, where you get as much attention as you can. “

Already have a lot of interest. Open-R 1 project made 10,000 stars in just three days of Githab. For stars GitHub users like this a project or a way to think it’s useful.

If the Open-R1 project is successful, AI researchers will be able to work on the top of the training pipeline and to work for the next generation’s Open source logic models, Bakch. He hopes that the Open-R 1 project will not only achieve a strong open source transcript of R1, but will also be a basis for better models.

“Open source development immediately benefits everyone, including the border lab and model provider, instead of being a zero-sum of games, as they can all use the same innovations,” Buckch said.

Although some AI experts have expressed concern over the possibility of open source AI abuse, Bakch believes that the benefits are higher than the risk.

“When the R1 recipe has been replicated, anyone who can rent some GPU can create their own variants with their own data, disconnect the technology everywhere,” he said. “We are really excited about the recent open source release, which is strengthening the role of openness in the AI. This is an important change for the field that changes the narrative that only a few handful of labs are able to make progress and that open source is lagging behind. “