Cohere ships Command A+ MoE

Cohere said on May 20 that it released Command A+, an open-source sparse Mixture-of-Experts model aimed at enterprise AI workloads that need lower infrastructure demands and tighter deployment control. The company described the model as its “fastest and most powerful” language model yet and said it was built for agentic, multilingual and multimodal tasks. Cohere said the model is available through its standard API endpoints and as open weights under an Apache 2.0 license. ### How big is Command A+, and what does “sparse” mean here? Cohere’s documentation says Command A+ has 218 billion total parameters but only 25 billion active parameters per token. The Hugging Face model card says the decoder-only sparse transformer uses 128 experts, with eight active per token, plus one shared expert applied to all tokens. That is the core of the company’s efficiency pitch: a larger total model with a smaller active compute path at inference time. (cohere.com) The same documentation says the model is designed for agentic, reasoning, vision and multilingual tasks. Cohere lists support for 48 languages in its launch materials and positions the model inside its Command family for tool-using agents and retrieval-augmented generation. ### Why is Cohere emphasizing two H100 GPUs? Cohere said Command A+ can run on as little as two Nvidia H100 GPUs. (docs.cohere.com) In its launch post, the company tied that claim to “sovereign agentic capabilities” and said the model was built for customers that need local control over data and infrastructure. A Business Wire release distributed by Cohere said the target users include governments and regulated industries. The Hugging Face release also includes quantized variants, including FP8, alongside the BF16 model. Cohere and secondary coverage said those formats are intended to reduce deployment costs and hardware requirements, although Cohere’s primary materials center the two-H100 claim rather than broader benchmark comparisons against rivals. ### What enterprise features is Cohere attaching to the launch? (cohere.com) Cohere said the model is built for complex reasoning, multimodal inputs and multilingual enterprise tasks. The company’s documentation says Command A+ supports vision inputs, and VentureBeat reported that Cohere also added native citation features and “lossless quantization” as part of the release. Cohere’s release notes say private deployment options are available for enterprise customers. (huggingface.co) The company framed those options around security and control, while the press release said the launch was aimed at “sovereign critical infrastructure” use cases. ### How does this fit with Cohere’s earlier Command models? Cohere released Command A in 2025 as a 111 billion-parameter model for tool use, agents and retrieval workloads, according to its documentation and Hugging Face model card. (cohere.com) The newer Command A+ keeps the enterprise and agentic positioning but adds a sparse MoE architecture, multimodal support and broader multilingual coverage. (docs.cohere.com) Cohere’s model overview page now lists Command A+, Command A, Command A Reasoning, Command A Vision and Command A Translate within the same family. That lineup suggests the company is packaging separate enterprise use cases around a shared Command stack rather than a single general-purpose release. ### Where can users get it now? Command A+ is already live in Cohere’s API as `command-a-plus-05-2026`, according to the company’s changelog. (docs.cohere.com) The open-weight release is also posted on Hugging Face in BF16 and compressed variants under an Apache 2.0 license. Cohere said users can find technical specifications and implementation details in its model documentation, while enterprise buyers can pursue private deployment through the company’s standard channels. (docs.cohere.com 1) (docs.cohere.com 2)