Mozilla Wants to Build the Real Open AI

Two decades after challenging Internet Explorer’s dominance with the Firefox browser, Mozilla now sees a similar opportunity to lead in artificial intelligence. Its new mission is to pry AI, piece by piece, from the control of a handful of large technology companies and push it back toward a more open ecosystem.

This time, the adversary is not a single corporation but a prevailing assumption—that the most powerful AI systems must remain locked behind proprietary walls. 

David vs Goliath

Today, only a small handful of companies control the most widely used AI models. While open models and frameworks have advanced rapidly, many startups and software companies continue to build their products on proprietary systems accessed through simple APIs.

The pattern was apparent at the India AI Impact Summit in New Delhi, where OpenAI announced partnerships with several major Indian companies—including MakeMyTrip, Zomato, JioHotstar, and ixigo—to integrate its models into consumer applications. It also signed an enterprise agreement with Tata Consultancy Services, enabling the deployment of its enterprise AI products across the firm and its workforce.

Many AI features on these consumer apps work on top of OpenAI’s proprietary models. While companies can customise how the systems behave within their products, the underlying models remain closed and controlled by the provider.

This trend persists even as open models improve rapidly, with some already rivalling proprietary ones in performance. Yet many developers still choose proprietary APIs.

Mozilla CTO Raffi Krikorian believes the issue narrows down to how usable open models are. “You could go download an Alibaba model—but then a question is, how do I use it?” Krikorian tells AIM in an interview.

He referred to the simplicity of integrating with closed-source models, where API calls make it straightforward for developers building AI-powered applications. “We need that experience, but for open source,” he said.

AI’s Rust Moment

Mozilla’s answer is a suite of developer tools, including any-llm, any-agent, any-guardrail, and several other open-source frameworks built under its Mozilla.ai initiative—launched in 2023 with an initial investment of $30 million from the company to built trustworthy, independent, and open-source AI ecosystem. The software community is attempting is to simplify the experience for developers working with open models by creating a common abstraction layer that allows different model providers to be accessed through the same interface.

The any-llm framework, for instance, lets developers switch between LLM providers with a simple configuration change rather than rewriting application code, allowing the same software to work with both proprietary and open models.

“This is important because we’re going to get to this place where certain AI models are good for certain things. We want developers to have that flexibility of wanting to rapidly test an open model and move to it,” Kirkorian explains.

Similarly, the any-agent framework provides a single interface for running and testing different agent frameworks. Developers can swap frameworks without rewriting their application logic. It supports popular systems such as LangChain, LlamaIndex, Google ADK, and others.

Another example is the llamafile framework, which allows developers to run LLMs locally via a single executable file.

This is remeniscent of how the organisation helped improve the software development experience with Rust as a modern alternative to C++. Rust is now not one of the most widely adopted programming languages in the ecosystem.

So is the organisation headed for another Rust moment, this time with Mozilla.ai?

“It’s hard to predict these things—but we’re going to keep taking shots on goal on these kinds of opportunities, and one of them will become the Rust-like moment,” Kirkorian remarks. “The thing about Rust was that it made it easier for people to write both type-safe and memory-safe code—you can also think about that as reducing the barriers to do the right things.”

He refers to the upcoming AI mode in the Firefox browser and how the organisation is considering implementing new features along these lines. “Can we start exposing APIs that allow you to build AI experiences in the browser? Because it’s Firefox, we can permission it properly along with the user context, and other things where developers can just start building.”

Mozilla’s Four-Pronged Strategy

Improving the developer experience is one of the four pillars through which Mozilla hopes to push the ecosystem towards openness. The other three are data, models, and compute.

Regarding data, Kirkorian says the current contract on the web for data harvesting is terrible, allowing all data to be used for training models. Mozilla wants to move towards a world where people and communities can create valuable datasets while having a say in how they are used and sharing in the value they generate.

In 2025, Mozilla established the Mozilla Data Collective (MDC) to create a community-controlled marketplace and governance system for AI training data. The initiative aims to build a two-sided marketplace where people and organisations can sell their datasets, with proper licensing and provenance tracing, so that both frontier model labs and smaller model providers can legally access and use the data for training.

He adds that the initiative will unlock large pools of data that remains inaccessible because communities are uncertain how the data will be used.

“A good example of this is that MDC is in conversations with almost every radio station on the planet, which sits on archives of voice data in localised languages. But most of them don’t put it out there on the internet,” he says.

He points out that communities often already possess large amounts of culturally relevant data. For example, the RMN broadcasting network in the Philippines has hundreds of hours of Tagalog audio available for model training. Making such datasets available, he argues, could enable the development of local AI systems that better reflect regional languages, values, and historical context rather than relying solely on globally dominant datasets.

“We can start to create a world where people can have models that actually represent them as opposed to a West Coast context in the United States,” he comments.

This would also strengthen the third pillar—models—by making smaller, domain-specific, and multilingual systems more accessible to developers.

Making AI Affordable

With regard to compute, Kirkorian says that more doors need to open through distributed computing, sovereign clouds, and the productive use of idle GPUs. “I would love to get to a world where we can train on heterogeneous hardware,” he remarks.

Another way Mozilla is pursuing this strategy is through investments in relevant startups. 

Mozilla Ventures participated in the $50 million seed round in Adoption Labs this year. 

Its founder, Sara Hooker, an AI researcher and advocate for cheaper AI systems, is attempting to replace the “bigger models + more GPUs” paradigm with smaller, adaptive AI systems that learn continuously and use compute more efficiently.

Mozilla Ventures has also invested in is Flower AI, which builds infrastructure that allows AI models to be trained across many distributed devices or organisations while keeping the underlying data private and local.

Its portfolio also includes Hugging Face, Cinder, and Answer.AI.

Currently, the Mozilla Foundation has $1.4 billion in reserves and no debt, with $620 million in asset growth since 2020. The organisation plans to use this capital to invest in such startups and build additional AI frameworks.

“Mozilla is never gonna train its own foundation model. With $1.4 billon in the bank, we’ll start partnering with people who look at the big questions of how to change the ecosystem.”

For Kirkorian, the long-term vision is a world with 7 billion AGI (artificial general intelligence) systems, instead of seven, referring to a future where developers have the freedom, ease, and control to build products and applications tailored to specific cultures, communities, and societies.

“I feel it is Mozilla’s job to enable those 7 billion AGIs to flourish and bloom.”