Cohere

Benchmarking: We used WMT26 benchmarks, which were released after we created the model, ensuring we couldn't train North Small Translate to its measurements. (5/6) https://t.co/amZsHNndd7