Maybe you think paying top dollar for proprietary AI models is still the only way to get serious work done. Well, a new report from Mozilla serves as a pretty clear reality check. The performance gap between closed US frontier AI models and the best open-weight ones coming out of China has shrunk to roughly 4.4 months.
That narrow window changes the math for tech teams. It forces a straightforward financial question: is a four-month technical head start really worth paying up to five times more per task?
Where closed models still hold the line
The numbers in Mozilla’s analysis show open models like Moonshot AI’s Kimi K3 and Z.ai’s GLM-5.2 matching closed heavyweights on major benchmarks. In neutral harness evaluations by Vals AI, GLM-5.2 landed within a single point of Anthropic’s Claude Opus models while running at less than a fifth of the cost per job.
Mozilla CTO Raffi Krikorian noted that closed models earn their markup in very specific territory. This applies mostly to expert-level work taking humans 8 to 12 hours to complete, alongside heavy retrieval tasks. For routine automation taking under eight hours, open models handle the workload just fine.
Companies aren’t treating this as an all-or-nothing choice anymore. DoorDash, for example, routes everyday tasks to open models like Kimi and saves premium systems like Claude Fable for complex, deadline-driven projects (via ArsTechnica).
The hidden catch with open weights
Downloading model weights gives developers freedom from vendor lock-in, but “open weight” isn’t the same thing as fully open source. Out of 16 notable releases in Mozilla’s report, none provided the full dataset and training code required by Open Source Initiative standards.
Then there is the hardware bill. Kimi K3 packs nearly 2.8 trillion total parameters, and running its 1.56TB checkpoint requires serious infrastructure—think clusters with 64 or more high-end accelerators. The weights are technically available, but unless a company has massive compute on hand, hosted APIs remain the only practical way to run them.
A shifting global balance
The report also highlights a sharp geopolitical divide. On gateway platform OpenRouter, eight of the top ten models by token volume are open-weight, and seven of those came from Chinese labs like DeepSeek, Alibaba, and Moonshot.
While US companies dominate proprietary closed AI, Chinese developers are effectively shaping the open ecosystem. To keep any single nation from setting global defaults, Mozilla is pushing Western institutions and public projects—like Switzerland’s Apertus compute program—to fund neutral, open reference models before the gap closes entirely.
The post Chinese Open Models Are Only 4 Months Behind Top US AI While Costing 5x Less appeared first on Android Headlines.
