How did your Chinchilla research on compute-optimal scaling shape Mistral architectural philosophy?

Chinchilla demonstrated that historical large models were severely under-trained relative to their computational budgets. By optimizing dataset token volume and model parameter ratios, we proved that smaller, meticulously trained architectures can match or exceed the performance of massive legacy systems. At Mistral, we focus relentlessly on computational efficiency, delivering frontier reasoning capabilities at a fraction of standard inference costs.