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