For decades, Mozilla has been the champion of an open internet. Now, the organization is making the same case for artificial intelligence. In a recent interview, Mozilla CTO Raffi Krikorian laid out a vision where open-source AI is treated not as a product to be sold, but as critical infrastructure to be protected and nurtured.
A narrowing gap between open and closed models
Krikorian points to compelling data from a Mozilla report showing that the performance gap between top open-source models and proprietary systems like Claude and ChatGPT has narrowed to just 3%. In February 2026, Alibaba's open-source model Qwen was downloaded more times than the next eight models combined. The open-source ecosystem is no longer a hobbyist space. It's a multi-hundred-billion-dollar commercial layer.
The debate over open AI has gained fresh momentum after recent events, but Krikorian believes businesses are already voting with their wallets. IT and HR teams are migrating to open models for price-performance reasons and control over data. Regulated industries want to self-host and keep data within their firewall. The shift is happening, even if it's not making headlines.
A policy categorization error
Krikorian's key argument is that the U.S. government is making a fundamental categorization error. It views intelligence as a product when it should be viewed as infrastructure. Products are things you rent or turn off. Infrastructure is foundational.
He sees the rest of the world wanting to buy AI as infrastructure, not a product. The U.S. government, however, is backing perceived winners, influenced by heavy lobbying from companies racing toward public listings. Krikorian expresses frustration at this "simple thing" approach, arguing it misses the bigger picture.
The China factor and the open-weight conversation
While open-source models are gaining popularity, Krikorian notes the popularity of Chinese models adds a layer of complexity. He suggests that if demand is high, other countries should be trying to fill it, rather than relying on one nation's offerings.
The conversation also touches on a crucial distinction: open source versus open weight. The strictest definition of open source requires understanding all the data used in training, evaluation systems, and the entire pre-training and post-training process. What many are using today are open-weight systems, which offer the ability to download, run, and fine-tune but lack full transparency.
Krikorian believes the path to truly local, representative AI lies in taking these open-weight models and fine-tuning them to reflect community values and languages. The technology still needs work, but the direction is clear.
The Firefox precedent
When asked why we should believe open-source AI won't end up like the web, dominated by a few companies, Krikorian points to Firefox's historical role. Even small market share has kept the web open. A vocal community of Firefox users has prevented Google from building a closed version of the internet.
He wants the same dynamic for AI. Enough open model traffic would force a conversation around model choice, user agency, and data privacy. And he argues that a large amount of traffic is already moving toward open models.
Open-source AI is a movement that aligns with the foundational principles of the internet. Krikorian's vision is for an AI internet where the "biggest provider" is a service company that helps others use and customize open intelligence. It's a future where AI serves everyone, not just the companies that build it.