Open-weight AI models are now only around four months behind the closed frontier, according to a new Mozilla report!

Kimi K3 nearly matches Sol on Terminal-Bench, while GLM-5.2 approaches Claude Opus performance at less than one-fifth the cost. Eight of the ten largest models on OpenRouter by token volume are now open-weight.

Closed models still lead on the hardest tasks, but the gap has become surprisingly small.