This person did a great comparison against Qwen models, and despite them having 8x less active params, they outperform the Cohere model in every category: https://x.com/DJLougen/status/2057196012918149368?s=20

I’m always surprised by how performant the Cohere models are. They output quick. I tested out the BF16 and it seems pretty good. I tried out the FP8 one and it did seem a bit dumber. Curious to see how this ranks in benchmarks

New open-source model from Cohere