Mistral CEO Critiques OpenAI’s Closed-Source Model, Champions Nvidia Partnership for Open-Source AI Revival
March 21, 2025 | Paris / Silicon Valley — Arthur Mensch, CEO of French AI innovator Mistral AI, has taken a firm stance against the increasingly closed nature of the AI ecosystem, blaming OpenAI’s 2020 release of GPT-3 for slowing down open-source AI momentum. Mensch delivered these remarks during a recent appearance on the Andreessen Horowitz (a16z) Podcast, alongside Nvidia CEO Jensen Huang.
OpenAI’s Shift to Proprietary Models Blamed for Disrupting AI Collaboration
Mensch argued that prior to the rise of proprietary models like GPT-3, global AI research was largely cooperative, with industry leaders like Google DeepMind, Meta AI, and Stanford AI Lab openly sharing breakthroughs. “Between 2010 and 2020, labs were building on top of each other’s innovations,” Mensch remarked. “That openness evaporated when OpenAI chose to withhold model weights and code.”
Following OpenAI’s pivot, other tech giants, including Google (via Gemini and PaLM 2) and Anthropic with Claude AI, adopted closed-source models. Today, most of the most powerful AI systems — such as GPT-4, Claude 3, and Gemini Ultra — remain inaccessible to the public.
Nvidia and Mistral Partnership Signals New Path for Open-Source AI
Despite this trend, Mistral AI is spearheading a counter-movement by advancing open-source alternatives. The company’s collaboration with Nvidia has already yielded NeMo, a highly competitive open-source large language model. “The goal is to return to the spirit of open AI research,” said Mensch. Nvidia’s backing of open initiatives has positioned it as a key player fostering innovation beyond Big Tech silos.
Open-Source Momentum Gains Global Allies
Alongside Mistral, other players are doubling down on open AI. Meta continues to iterate on its LLaMA series, most recently with LLaMA 3, while Alibaba’s Qwen models and DeepSeek’s R1 have demonstrated the strength of collaborative development.
Interestingly, DeepSeek R1, which recently disrupted ChatGPT’s dominance in app store charts, was built upon Alibaba’s Qwen, which itself borrowed heavily from Meta’s LLaMA. “It’s a positive feedback loop where everyone benefits,” Mensch added, highlighting how open-source frameworks allow smaller teams to punch above their weight.
OpenAI Defends Its Strategy Amid Growing Scrutiny
While OpenAI has yet to respond directly to Mensch’s criticism, the company has consistently argued that keeping its models proprietary is necessary for safety, citing risks like AI misuse and misinformation. Sam Altman, CEO of OpenAI, has defended the closed approach, especially following the release of GPT-4 Turbo and the specialized ChatGPT Gov model for U.S. government use.
The Bigger Picture: AI’s Fragmentation
The open vs. closed-source debate reflects deeper divides within the AI industry, pitting startup disruptors like Mistral, Hugging Face, and Cohere against established giants like Microsoft, Amazon Web Services (AWS), and Google Cloud. While AWS and Microsoft Azure host many proprietary models, open alternatives increasingly attract academic researchers and smaller enterprises.
Mensch’s remarks come as AI ethics bodies and government agencies, including the European Commission and U.S. Federal Trade Commission (FTC), weigh in on the risks of concentrated AI power in the hands of a few corporations.
What’s Next?
With Mistral AI preparing to release new iterations of its open models in collaboration with Nvidia, and Meta’s LLaMA 3 expected to hit public repositories soon, the open-source resurgence appears to be gathering momentum.
Leave a comment