Open-weight AI models close gap to frontier offerings, but deployment hurdles remain

Mozilla's latest State of Open Source AI report shows that leading Chinese open-weight AI models are now within about four months of frontier US models on task capability, and at a fraction of the cost. But the gap is more complicated than a single number suggests, and deployment realities may shift the equation for builders.

Mozilla report: Chinese open-weight models within months of frontier capability

Mozilla published version 1.1 of its State of Open Source AI report on Sept. 15 using data through Sept. 1, finding that several top Chinese open-weight AI models are closing the gap with US frontier offerings. On the Artificial Analysis Intelligence Index, the best open model trailed the closed leader by three points at 60% of the price, and was two points behind Claude Fable 5 at 30% of the price. Mozilla's fit on METR task-horizon data puts the open-closed gap at around 4.4 months, consistent with Epoch AI's four-month estimate. Open-weight models in this context mean downloadable weights, not training data or code. The report counts 16 notable open releases, but none meets the Open Source Initiative's definition.

OpenRouter usage: open weights dominate traffic, closed capture revenue

On OpenRouter, a marketplace routing developer traffic to hundreds of models, eight of the top ten models by August token volume were open weights, seven of them Chinese-built. Yet closed providers captured 96% of model-layer revenue on OpenRouter from May to September 2025, according to the Linux Foundation. Mozilla CTO Raffi Krikorian said the decision to pay for closed models appears workload-specific rather than organization-specific.

The hardware reality: open weights are not always deployable

The four-month gap and price comparisons are measured API to API on hosted endpoints at list price. When hardware constraints enter the picture, the gap widens. The best open model that fits one server scored 52.6 on Mozilla's hardware chart, and the best on one GPU scored 40, compared to the top closed model at 63. That is a 10- and 23-point drop, respectively, a larger gap than the reported four months.

Kimi K3, a leading Chinese open-weight model, has a native MXFP4 checkpoint of about 1.56TB across 96 shards. Mozilla's serving configuration lists 64 or more accelerators, and vLLM calls for at least eight GB300 GPUs with multiple nodes for production traffic. The report describes this as open but not runnable by most who hold it. One exception is Thinking Machines' Inkling-Small model, whose NVFP4 version fits one B300 at a 180GB floor.

What the narrowing gap means for builders

The implication is not that frontier APIs are obsolete, but that enterprises have more reason to evaluate open-weight options for workloads where cost or control matter. Krikorian's observation that premium model purchases are workload-driven rather than org-wide suggests teams should benchmark per-task rather than defaulting to a single provider. The report's data stops at Sept. 1. Since then, Artificial Analysis moved to index v4.3, and the live board shows Claude Fable 5.1 at 53 with Kimi K3 at 44, not directly comparable. Mozilla's own chart caption reads: "the gap resets every release cycle."

K3 also carries an unproven allegation in a Sept. 8 NSA/CISA/FBI joint advisory that Moonshot extracted Claude Fable 5 data through distillation. Mozilla's report calls the claim "asserted, and unshown."

For teams building AI products, the takeaway is to benchmark on your own workloads rather than relying on aggregate indices, and to model total cost including infrastructure, not just token price. The narrowing capability gap makes open-weight models a serious option, but only if you can run them.