Command A+ Runs a 218B Model on Two H100s Under Apache 2.0

Cohere released Command A+ on 20 May 2026: a 218-billion-parameter sparse mixture-of-experts model with 25 billion active parameters, under Apache 2.0. Weights are on Hugging Face as CohereLabs/command-a-plus-05-2026.

It runs on as few as two H100 GPUs, supports 48 languages and a 128K context window, and ships with native citation grounding. Cohere positions it for sovereign and on-premises enterprise agentic workflows - the deployment profile is the product, as much as the benchmark numbers.

The two-GPU requirement is the number that matters for anyone actually planning a deployment. A 218B model that fits on two H100s is within reach of a single server, which is a different procurement conversation from a model needing a multi-node cluster.

Cohere separately updated its enterprise RAG line with Command R+ 2026, extending the context window to 256,000 tokens with more reliable multi-step tool use and JSON output.

Command A+ is on our roster with its specification table populated. This item records the licence and hardware footprint, which are the facts most relevant to evaluating it against the other self-hostable models we cover.