Nvidia ropes in Sarvam, Mistral, Perplexity in global ‘Nemotron Coalition’ to build open AI models

In a major push toward open and sovereign AI, US chip giant Nvidia has brought together a group of global AI labs, including India’s Sarvam, to co-develop next-generation open foundation models under a new initiative called the Nemotron Coalition.

The coalition, announced at Nvidia's GTC conference, positions Sarvam among a select group of frontier AI players such as Mistral AI, Perplexity, Thinking Machines Lab, LangChain, Cursor, Reflection AI and Black Forest Labs, working jointly on building open models that can rival proprietary systems.

Sarvam’s India-first play at the global AI table

Sarvam will contribute its expertise in building multilingual and voice-first models tailored to local contexts, a key gap in current global AI systems.

“AI reaches its full potential when it works in every language and for every community,” said Pratyush Kumar, cofounder and CEO of Sarvam. “Open models make this possible by giving builders the freedom to adapt frontier capabilities to real-world needs.”

He added that Sarvam will help build models that are “voice first, language inclusive, understand local culture and provide a platform for developers to build applications that matter at population scale.”

For India, Sarvam’s inclusion signals a growing role in shaping global AI infrastructure, especially in areas like multilingual AI, voice interfaces and population-scale deployments.

The coalition’s open model approach also aligns with India’s broader push for sovereign AI under initiatives like the IndiaAI Mission, where access to foundational models, rather than reliance on foreign closed systems is seen as critical.

By participating in the Nemotron Coalition, Sarvam effectively places India at the table where next-generation foundation models are being built and not just consumed.

What is the Nemotron Coalition?

The Nemotron Coalition is being pitched as a first-of-its-kind global collaboration where AI labs will pool data, compute, evaluation frameworks and research expertise to build frontier models in an open ecosystem.

“Open models are the lifeblood of innovation and the engine of global participation in the AI revolution,” said Jensen Huang, founder and CEO of Nvidia. “The Nvidia Nemotron Coalition unites world-class AI labs to develop frontier open models that champion transparency, collaboration and sovereignty.”

The first model under the initiative will be co-developed by NVIDIA and French AI firm Mistral AI, and trained on NVIDIA’s DGX Cloud infrastructure. It will eventually form the foundation of the upcoming Nemotron 4 family of open models.

The move comes amid growing global debate around open vs proprietary AI systems, with companies like OpenAI, Google and Anthropic largely keeping their most advanced models closed.

By contrast, NVIDIA’s coalition aims to create a shared base model that developers and enterprises can customise for specific industries, geographies and use cases.

“Open frontier models are how AI becomes a true platform,” said Arthur Mensch, CEO of Mistral AI, which will play a leading role in training the base model.

Perplexity CEO Aravind Srinivas added that open models “give builders the flexibility to improve performance, reduce costs and push AI applications into everyday use.”

Building the “AI agent” era

Beyond raw intelligence, coalition members are also focusing on making models more usable in real-world applications, especially agentic AI systems.

“With over 100 million monthly downloads of LangChain’s frameworks, we’ve seen that frontier models must go beyond raw intelligence to enable reliable tool use, long-horizon reasoning and agent coordination,” said Harrison Chase, CEO of LangChain.

LangChain will contribute agent frameworks and evaluation tools, while Cursor will bring real-world performance datasets, and Black Forest Labs will add multimodal capabilities spanning images and video.

Who brings what to the table?

Each member is expected to contribute specialised capabilities:

• Black Forest Labs: multimodal AI including images, video and action prediction
   
• Cursor: real-world evaluation datasets and performance benchmarks
   
• LangChain: agent frameworks, tool use and long-horizon reasoning
   
• Mistral AI: efficient, customisable frontier model development
   
• Perplexity: high-performing, accessible AI systems
   
• Reflection AI: dependable and safe open AI systems
   
• Sarvam AI: sovereign language AI and multilingual capabilities
   
• Thinking Machines Lab: data collaboration and research support