Editor's note · week of Apr 21
Gemini 2.5 Flash continues to dominate the translation category while smaller open models from Meta and Qwen close the gap on reasoning tasks. The most interesting development this week: Cerebras inference speeds have surpassed 1000 tok/s on some models — redefining what "fast" means for real-time apps.