25B active / 218B total (MoE)
128K input / 64K output
Apache 2.0
Official open-weight Command A family release that Cohere positions as its fastest and most powerful language model yet, combining text, vision input, reasoning, translation, and agentic tool-use capabilities in one model. The Hugging Face model card describes Command A+ as a decoder-only sparse Mixture-of-Experts Transformer with 25B active parameters and 218B total parameters, using 128 experts with 8 active per token plus a shared expert. Cohere says the architecture interleaves sliding-window RoPE attention and global attention in a 3:1 ratio inherited from Command A, uses a fully dropless sparse MoE layer with token-choice routing, additive-bias load balancing, and a normalized-sigmoid top-k router. Release materials emphasize 48-language coverage, stronger agentic and tool-use performance than earlier Command-family models, and deployment across multiple official quantizations: BF16, FP8, and W4A4. The W4A4 card specifically highlights quantization-aware distillation and selective NVFP4 W4A4 quantization of MoE experts while keeping the attention path at full precision, targeting single-B200 or dual-H100 deployment with lower latency and higher throughput.