The performance gap between frontier AI models from US companies and the best open-weight models from China has narrowed to just 4.4 months, according to a Mozilla report published on September 15 and obtained by Ars Technica. The study, titled State of Open Source AI, shows that US leadership has never been so short.
The report highlights that the Kimi K3 model, from Chinese startup Moonshot AI, achieved competitive performance with the most advanced Silicon Valley models. The training cost difference underscores the achievement: while closed frontier models cost about five times more to train, Kimi K3 was trained for a fraction of that amount.
For organizations deciding which model to adopt in production, the report recommends open-weight models as the default option for most tasks. The rationale is that the performance gap no longer justifies the extra cost and vendor lock-in in most usage scenarios.
Mozilla does not claim that open models outperform closed ones across all tasks. The 4.4-month advantage means that for applications that do not require absolute state-of-the-art, the latest open model delivers quality equivalent to a frontier model from less than six months ago, at one-fifth the training cost.
This publication considers that the Mozilla report provides a concrete basis for infrastructure teams to reassess API contracts with closed-model providers. Instead of paying a premium for theoretical leadership, operations that can tolerate a four-month lag can drastically reduce inference costs without significant quality loss in results.