Mozilla Report: China’s Open-Weight Models Narrow Gap With US Frontier to 4.4 Months

Mozilla released the second edition of its State of Open Source AI report on September 15, 2026, saying the capability gap between China’s best open-weight models and leading closed models from US tech companies has narrowed to 4.4 months.

Compared with the first report published in July 2026, the new edition focuses on specific models, task duration and per-task cost. The report highlights Moonshot AI’s Kimi K3, which scored only 3 points lower than Anthropic’s Fable 5 on the Artificial Analysis Intelligence Index while costing about 30% as much.

On capability, METR measures the “time horizon” of a model by how long it can reliably complete tasks, using a 50% success threshold. The best closed model can handle tasks of roughly 12 hours, 1.7 times the about 7 hours achieved by the best open model. Mozilla CTO Raffi Krikorian said the difference is task-dependent: the premium for closed models mainly appears in specialized tasks that take 8 to 12 hours, while open models can handle routine work at very low cost. He said that by January 2027, open models will handle about 12-hour tasks and closed models roughly 20-hour tasks.

The shift is already affecting enterprise usage. DoorDash uses Kimi for routine tasks and keeps Fable for more complex work. On the routing platform OpenRouter, 8 of the top 10 models by token usage in August 2026 were open-source models.