Cohere ships Command A+

Cohere released Command A+ on May 20, 2026, describing it as its fastest and most powerful language model and the last model in the Command A family. The company said the model is open-source, built for enterprise use, and combines reasoning, vision, multilingual and agentic capabilities in one system. Cohere’s documentation says Command A+ uses a sparse Mixture-of-Experts architecture with 218 billion total parameters and 25 billion active parameters. Cohere said that design allows the model to run on as little as two Nvidia H100 GPUs. ### What exactly did Cohere ship? Command A+ is a decoder-only sparse MoE model that Cohere positioned as a single replacement for several earlier Command variants focused on separate tasks. In its release notes, Cohere said the model combines support for vision inputs, reasoning, translation and agentic tasks “within the same model.” (cohere.com) Cohere’s model page says Command A+ is aimed at agentic, reasoning, vision and multilingual tasks. The company lists 49 supported languages, a 256,000-token context window and availability through its chat endpoints under the model ID `command-a-plus-05-2026`. ### Why does the two-H100 claim matter? Cohere said the model’s sparse architecture activates only 25 billion parameters per token out of 218 billion total. (docs.cohere.com) That is the core of the company’s efficiency pitch: a model large enough to market as frontier-class, but with a serving footprint closer to a much smaller dense model. Microsoft, which said on May 21 that Command A+ is available in Microsoft Foundry, repeated Cohere’s description of the model as a 218B MoE system with 25B active parameters and said efficiency was a core design goal. (docs.cohere.com) Microsoft’s post said Cohere claims the model can run on a single Nvidia Blackwell GPU or two H100 GPUs. ### Is this different from Cohere’s earlier Command models? Command A, the earlier flagship in the family, was a 111-billion-parameter model that Cohere said could also run on two A100 or H100 GPUs. Cohere described that model as strong on enterprise tasks including tool use, retrieval-augmented generation and agents. Command A+ extends that line by adding sparse MoE routing and folding multiple specialized capabilities into one release. (techcommunity.microsoft.com) Cohere’s release notes call it the company’s first Mixture-of-Experts model, while the model page says it is designed to handle reasoning, multimodal and multilingual workloads in one system. ### Where is Cohere aiming this model? (docs.cohere.com) Cohere’s blog framed the release around “sovereign agentic capabilities,” a phrase the company used to signal customers that want to run advanced models with tighter control over infrastructure and deployment. The company called Command A+ an “enterprise workhorse” for complex reasoning, multimodal and multilingual agentic tasks. (docs.cohere.com) The emphasis on open-source distribution and lower hardware requirements puts the model in a part of the market where enterprises and governments are weighing performance against cost, control and where workloads run. Cohere’s public materials focus on private deployment, customization and enterprise automation rather than consumer chat use. ### What should readers watch next? (cohere.com) May 21 brought an immediate distribution step: Microsoft said Command A+ is now available in Microsoft Foundry. Cohere’s documentation and release notes remain the primary places to track model specifications, endpoint support and any updates to deployment guidance. (techcommunity.microsoft.com) (cohere.com)