DeepSeek breakthrough emboldens open-source AI models like Meta Llama

DeepSeek’s powerful new artificial intelligence model isn’t just a win for China — it’s a victory for open-source versions of the tech from the likes of Meta, Databricks, Mistral and Hugging Face, according to industry experts who spoke with CNBC.

Last month, DeepSeek released R1, an open-source reasoning model that claims to rival the performance of OpenAI’s o1 model using a cheaper, less energy-intensive process.

The development caused the market values of Nvidia and other chipmakers to plummet on fears that it could lead to reduced spending on high-performance computing infrastructure.

DeepSeek is a Chinese AI lab that focuses on developing large language models with the ultimate aim of achieving artificial general intelligence, or AGI. It was founded in 2023 by Liang Wenfeng, co-founder of AI-focused quantitative hedge fund High-Flyer.

AGI loosely refers to the idea of an AI that equals or surpasses human intellect on a wide range of tasks.

Since OpenAI’s ChatGPT burst onto the scene in November 2022, AI researchers have been working hard to understand and improve upon the advances of the foundational large language model technology that underpins it.

One area of focus for many labs has been open-source AI. Open source refers to software whose source code is made freely available on the open web for possible modification and redistribution.

Plenty of firms from tech giants like Meta to scrappier startups such as Mistral and Hugging Face have been betting on open-source as a way to improve on the technology while also sharing important developments with the wider research community.

DeepSeek’s technological breakthrough has only made the case for open-source AI models stronger, according to some tech executives.

Seena Rejal, chief commercial officer of AI startup NetMind, told CNBC the Chinese firm’s success shows that open-source AI is “no longer just a non commercial research initiative but a viable, scalable alternative to closed models” like OpenAI’s GPT.

“DeepSeek R1 has demonstrated that open-source models can achieve state-of-the-art performance, rivaling proprietary models from OpenAI and others,” Rejal told CNBC. “This challenges the belief that only closed-source models can dominate innovation in this space.”

Rejal isn’t alone. Yann LeCun, Meta’s chief AI scientist, said DeepSeek’s success represents a victory for open-source AI models, not necessarily a win for China over the United States. Meta is behind a popular open-source AI model called Llama.

“To people who see the performance of DeepSeek and think: ‘China is surpassing the U.S. in AI.’ You are reading this wrong. The correct reading is: ‘Open source models are surpassing proprietary ones’,” he said in a post on LinkedIn.

“DeepSeek has profited from open research and open source (e.g. PyTorch and Llama from Meta). They came up with new ideas and built them on top of other people’s work. Because their work is published and open source, everyone can profit from it. That is the power of open research and open source.”

Cut off by Washington from accessing advanced chips needed to train and run AI models, China has turned to open-source technology to boost the appeal of its AI models. Many Chinese firms — DeepSeek included — are pursuing open source models as a way to increase innovation and spread their use.

But the trend of companies turning to open-source technologies for success in AI isn’t limited to China. In Europe, an alliance of academics, companies and data centers have partnered on developing a family of high-performing, multilingual large language models, called OpenEuroLLM.

The alliance is led by Jan Hajič, a renowned computational linguist at Charles University, Czechia, and Peter Sarlin, the co-founder of Silo AI, an AI lab that was bought by U.S. chipmaker AMD last year.

The initiative forms part of a broader push for “AI sovereignty,” in which countries are encouraging investment in their own domestic AI labs and data centers to reduce a reliance on Silicon Valley.

There are downsides to open-source AI, however. Experts warn that, although open-source tech is a good thing for innovation, it is also more prone to cyber exploitation. That’s because it can be repackaged and modified by anyone.

Cybersecurity firms have already discovered vulnerabilities in DeepSeek’s AI models. Research that Cisco released last week revealed that R1 contained critical safety flaws.

Using “algorithmic jailbreaking techniques,” Cisco’s AI safety research team says it got R1 to provide affirmative responses to a series of harmful prompts from the popular HarmBench “with a 100% attack success rate.”

“DeepSeek R1 was purportedly trained with a fraction of the budgets that other frontier model providers spend on developing their models. However, it comes at a different cost: safety and security,” Cisco researchers Paul Kassianik and Amin Karbasi wrote.

Data leakage is also a concern. Data processed by DeepSeek’s R1 model via its website or app is sent straight to China. Chinese tech firms have long been dogged by allegations that Beijing uses their systems to spy on Western entities and individuals.

“DeepSeek, like other generative AI platforms, presents a double-edged sword for businesses and individuals alike,” said Matt Cooke, cybersecurity strategist EMEA at Proofpoint. “While the potential for innovation is undeniable, the risk of data leakage is a serious concern.”

“DeepSeek is relatively new, and it will take time to learn about the technology; however, what we do know is feeding sensitive company data or personal information into these systems is like handing attackers a loaded weapon,” Cooke added.

NetMind’s Rejal told CNBC that open-source AI models introduce cybersecurity risks which businesses need to consider, including software supply chain attacks, prompt jailbreaking and so-called “data poisoning” events that try to introduce biases or harmful outputs.

WATCH: Why China’s DeepSeek is putting America’s AI lead in jeopardy