The latest Data Phoenix news tagged Command A.
Cohere detailed a hardware-aware dynamic speculative-decoding technique it says runs about 23% faster than fixed-K decoding on its Command A model.