Mozilla Report: Open AI Models Narrow Performance Gap With Closed Frontier Systems

A new report from Mozilla finds that the performance gap between closed frontier AI models from US tech companies and the best open-weights models from Chinese developers has shrunk considerably.

The State of Open Source AI report, published September 15, says a leading open model now trails a top closed frontier model by a small margin on a composite intelligence benchmark while costing a fraction of the price.

According to Mozilla CTO Raffi Krikorian, paying for closed models makes sense for specific workloads rather than entire organizations. He pointed to expert professional work, high-intensity retrieval, and long context as areas where closed models justify their premium.

Open-weights models let anyone download the main components and run them locally, though developers typically withhold training data, pipelines, and code. Closed models from companies like Anthropic and OpenAI remain fully proprietary.

Krikorian said organizations still choose closed models because they work out of the box and include compliance packaging, support, and accountability. Many organizations, he noted, lack the staff to run open-weights models effectively.

Source: Ars Technica