The Apriel-1.6-15B-Thinker model newly released by the ServiceNow research team is truly impressive—it achieves performance on par with models in the same class, but with just 15B parameters and a model size reduced by a factor of 15. Even more impressive, 15% of its pre-training data was powered by NVIDIA, making its computational optimization approach worth noting.

If the path of high performance with smaller models proves successful, it could significantly impact the entire AI computing power market landscape. After all, everyone is currently burning money to stack parameters, so a solution that can do the same job with fewer resources is certainly more attractive.