According to Mozilla’s latest State of Open Source AI report, the performance gap between top proprietary US AI models and open-weights models has narrowed to 4.4 months. The report highlights that open models like Moonshot AI’s Kimi K3 deliver performance close to top closed models like Anthropic’s Fable 5 while operating at approximately 30 percent of the cost.
Mozilla’s analysis indicates that closed frontier models retain a distinct advantage primarily in complex, multi-step tasks requiring long contexts and expert-level reasoning. Based on task time horizons evaluated by METR, closed models reliably handle tasks lasting up to 12 hours of human expert work, whereas current open models handle tasks up to seven hours. However, open models catch up to those capability levels within roughly four months.
Due to this shrinking gap, enterprises are increasingly adopting a hybrid strategy. Companies like DoorDash report routing routine, shorter-duration workloads to cost-effective open models while reserving expensive closed frontier models for high-complexity tasks.
Why it matters
Encourages enterprises to adopt hybrid model architectures, routing routine workloads to open models to cut costs by 70%.
Shows open-weights models rapidly closing the capability gap, limiting closed frontier models to highly complex, long-horizon tasks.
Provides clear operational guidance on choosing models based on task duration and expert complexity.
Source: arstechnica.com
