On September 17, Mozilla released the second edition of its "State of Open Source AI" report, finding that the capability gap between China's top open-weight AI models and leading US closed-source models has narrowed to 4.4 months.
The report highlighted Moonshot AI's Kimi K3 open-weight model, which scored only 3 points below Anthropic's closed-source Fable 5 on the Artificial Analysis intelligence index while costing about 30% as much to run—showcasing strong performance-per-cost. On long-horizon tasks, the best open model can reliably handle about 7-hour tasks versus roughly 12 hours for top closed models.
Mozilla CTO Raffi Krikorian said open and closed models each suit different scenarios; at the current iteration pace, open models will handle 12-hour tasks within about four months, by which time closed models may reach 20-hour capability. Overseas firm DoorDash has adopted Kimi for daily operations while reserving Fable for complex work.
In August 2026, eight of the top ten models by token calls on OpenRouter were open-source, underscoring the expanding industrial deployment of open AI. The report also cautioned distinguishing "open source" from "open weight," as many weight-download products still carry restrictions.
Source: Global Times / Mozilla
