Sparse MoE architecture, 25B active parameters save computing power down to the core.

Cohere officially open-sourced the 218 billion parameter sparse mixture of experts model Command A+, licensed under Apache 2.0, targeting enterprise-level agents and private deployment, emphasizing data sovereignty and physical isolation. The full 218B model activates 25B tokens per inference; it can run on two H100s or a single B200, with low-precision versions like W4A4 available from Hugging Face. Command A+ natively supports multimodal inputs, with a 128K input context and 64K output length, designed for complex reasoning, autonomous tool invocation, database queries, and workflows involving long documents, supporting 48 languages (including EU official languages).