A new report from Mozilla, based on data from 1,400 developers and OpenRouter traffic, suggests that open-weight AI models now trail top US-based closed models by approximately four months. The study indicates that open-source capabilities are doubling every 3.9 months, compared to 5.5 months for closed systems.
Performance benchmarks show that models like Z.ai’s GLM-5.2 perform close to Claude Opus 4.7 while costing less than one-fifth per test. Despite this, closed-source providers captured 96% of model-layer revenue on the OpenRouter marketplace between May and September 2025.
The report highlights significant hardware requirements for running top open models. For instance, the Kimi K3 model requires approximately 1.56TB of memory across 96 shards, necessitating clusters of 64 or more accelerators. Most open models remain difficult for individual users to host locally.
Mozilla notes that the four-month performance gap is not static and resets with each new release cycle. Additionally, the report mentions allegations from a joint NSA, CISA, and FBI advisory regarding the potential use of data distillation from Claude Fable 5 to train the Kimi K3 model.
