Cohere Releases Command A+: A 218B Sparse MoE Model for Agentic Workflows That Runs on as Few as Two H100 GPUs

Cohere released Command A+, an Apache 2.0 open-source sparse MoE Transformer for enterprise AI.

The model has 218B total parameters with 25B active, 128K context, 64K generation, and supports text, image, and tool use.

Command A+ combines reasoning, vision, translation, and agentic capabilities into one model, with BF16, FP8, and recommended W4A4 variants.

Cohere says it improves benchmark scores and latency over prior Command A models across coding, QA, spreadsheets, memory, and multimodal tasks, while expanding multilingual coverage.