The race to build the most powerful AI models has largely been a competition, with labs guarding their research, their data, and their techniques closely. NVIDIA is trying something different. The company has announced the NVIDIA Nemotron Coalition, a global collaboration between some of the most respected names in open AI development, aimed at building frontier-level foundation models through shared expertise, shared data, and shared compute.
The inaugural members are Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam, and Thinking Machines Lab. Together, they represent a broad cross-section of the AI ecosystem, from multimodal generation and autonomous agents to sovereign language AI and frontier model research. The first model built by the coalition will be codeveloped by NVIDIA and Mistral AI, trained on NVIDIA DGX Cloud, and released as open source. It will form the foundation of the upcoming NVIDIA Nemotron 4 family of models.
Why This Coalition Exists
The premise behind the Nemotron Coalition is straightforward but meaningful. Building frontier AI models requires enormous resources: vast amounts of high-quality training data, significant compute infrastructure, and deep expertise across many different domains. No single organization has all of these things in equal measure. By pooling contributions from multiple specialized organizations, the coalition aims to build something more capable and more broadly useful than any member could produce independently.
The open source dimension is central to the project’s purpose. Once the model is built, it will be made available to developers and organizations worldwide to post-train, specialize, and adapt for their specific industries, regions, and needs. That downstream flexibility is what distinguishes open frontier models from proprietary ones, and it is what allows a single shared foundation to support a wide variety of applications across many different contexts.
Jensen Huang, founder and CEO of NVIDIA, framed the coalition’s mission in terms of global participation, describing open models as the engine that allows students, scientists, startups, and entire industries to participate in the AI revolution. He emphasized the values of transparency, collaboration, and sovereignty as the principles the coalition is designed to embody.
What Each Member Brings to the Table
The coalition is designed around complementary contributions rather than duplicated effort, with each member contributing the specific expertise and resources it is best positioned to provide.
Black Forest Labs, known for its work on multimodal generative models covering images, real-time video, and action prediction, will contribute its expertise in visual intelligence. Founder and CEO Robin Rombach noted that open models have always been central to driving frontier capabilities, and that coalitions between independent partners make it possible to reach the scale needed to advance state-of-the-art multimodal systems.
Cursor, the AI-native software development company, will contribute real-world performance requirements and evaluation datasets to improve the quality and reliability of the base model for developers. Cofounder and CEO Michael Truell pointed out that when frontier models are accessible and transparent, developers everywhere can help shape how the technology evolves.
LangChain, whose frameworks are downloaded more than 100 million times per month, will focus on the agent harness for the models, rigorously evaluating agent capabilities and providing observability into agent behavior. Cofounder and CEO Harrison Chase noted that frontier models must go beyond raw intelligence to enable reliable tool use, long-horizon reasoning, and agent coordination, and that LangChain’s contribution will help make Nemotron models the best foundation for the next generation of AI agents.
Mistral AI is taking a leading role in the actual training of the first coalition model, codeveloping it alongside NVIDIA. The company brings its well-established expertise in building efficient, customizable models that offer developers full control over how they are used and adapted. Cofounder and CEO Arthur Mensch described open frontier models as how AI becomes a true platform, and said the goal is to help establish a global foundation for AI that empowers developers to build the next generation of applications.
Perplexity will contribute its frontier model development expertise with a focus on making AI accessible at scale. Cofounder and CEO Aravind Srinivas emphasized the practical value of open models in giving builders the flexibility to improve performance, reduce costs, and push AI applications into everyday use.
Reflection AI is joining the coalition with a focus on building dependable and safe open systems. Cofounder and CEO Misha Laskin described the coalition as part of a broader effort to ensure that the foundation of AI intelligence remains open and accessible worldwide rather than controlled by a small number of organizations.
Sarvam will contribute its expertise in sovereign language AI development, with a particular focus on voice-first, language-inclusive AI that works across different languages and local cultures. Cofounder and CEO Pratyush Kumar noted that AI reaches its full potential when it works for every language and every community, and that open models make that possible by giving builders the freedom to adapt frontier capabilities to real-world needs.
Thinking Machines Lab, founded by Mira Murati, will contribute data collaboration and research support. Murati described the coalition’s mission as aligned with the goals behind the company’s Tinker platform, which was built around the principles of adaptability, collaboration, and broad accessibility.
The First Model and What Comes Next
The first tangible output from the coalition will be a base model codeveloped by NVIDIA and Mistral AI and trained on NVIDIA DGX Cloud. The other coalition members will contribute data, evaluation frameworks, and domain expertise to support the model’s post-training and continued development. Once complete, the model will be open sourced and will serve as the foundation for the Nemotron 4 family.
The Nemotron 4 family is positioned as the next major step in NVIDIA’s open model roadmap, building on the Nemotron 3 models announced separately at GTC this week. With Mistral AI’s model-building expertise and NVIDIA’s compute infrastructure at the center of the training process, and the broader coalition contributing specialized knowledge across agents, multimodal systems, safety, and language inclusivity, the resulting model is designed to be more capable and more broadly applicable than either organization could build alone.
What This Means for the Open AI Ecosystem
The Nemotron Coalition is notable not just for what it will produce but for what it represents as an organizational model. The history of frontier AI development has been characterized by intense competition and significant secrecy, with the most capable models typically kept proprietary by the organizations that build them. Coalition-based development of open frontier models is a meaningful departure from that pattern.
If the model is genuinely good and genuinely open, it creates a shared foundation that any developer or organization can build on, which tends to accelerate innovation across the ecosystem rather than concentrating it in a few places. The diversity of the coalition members, spanning everything from a startup focused on Indian language AI to a company building visual generation models to one of Europe’s most prominent frontier model labs, suggests that the shared foundation will have real breadth.
Whether a collaborative approach to frontier model development can keep pace with the proprietary models being built by the largest AI labs is an open question. But the scale of compute NVIDIA can bring to the training process, combined with the specialized expertise of eight well-regarded coalition members, makes this a serious effort rather than a symbolic one. The first model will be the real test of what this approach can produce.