Cohere has open-sourced Command A+, a 218-billion-parameter sparse mixture-of-experts (MoE) model under the Apache 2.0 license. It's built for enterprise agents and private deployment, with a focus on data sovereignty — meaning companies can run it on their own hardware without any vendor lock-in.
Though the model has 218 billion parameters total, it's sparse — only about 25 billion are activated per forward pass. That efficiency trick means you can run it on as few as two NVIDIA H100s or a single B200 GPU. Hugging Face already hosts quantized versions like W4A4 to make deployment even cheaper.
Command A+ handles both text and images natively, with a 128,000-token input context and 64,000-token output. Cohere says it's optimized for agentic tasks — think complex reasoning, tool calling, and database queries — along with long-document processing. It works in 48 languages, including every official EU language.