Mozilla's inaugural assessment of open source AI in 2026 includes the finding that there's near-parity on capability and that open source models account for a third of all tokens. It warns that the real contest has moved up a layer to the harness.

The State of Open Source AI 2026 is Mozilla's first annual assessment of the open source AI landscape, drawing on a SlashData developer survey, OpenRouter data, the Linux Foundation, and public filings. Its headline finding is that open models have effectively caught up. The average capability gap to the top closed models is down to 3.3% (albeit up from 0% a little while ago). Inference costs have fallen roughly 50× in three years And open weights now serve about a third of all tokens routed on OpenRouter. It's also become a real business, with companies commanding multi-billion-dollar valuations. The report's central argument is that the fight has moved off the model itself. Open source models ships easily but are hard to deploy. The gap is operational tooling, standardization, and enterprise readiness, not raw capability.