Cohere Unveils Command A+, Advancing Sovereign And Efficient Enterprise AI Deployment

Enterprise AI company Cohere has announced the release of Command A+, an open-source mixture-of-experts language model designed for enterprise-grade agentic tasks with relatively low compute requirements.

The company said the model builds on a year of work inside North, Cohere’s enterprise workspace for developing and deploying agentic AI systems, and brings together capabilities from across the Command family into a single model intended for scalable use across experimentation, deployment, and production environments.

Command A+ is now available under the Apache 2.0 license, giving developers broad access to a model Cohere says is intended to support sovereign AI deployment, including environments where organizations want direct control over infrastructure and model behavior. The release includes downloadable weights on Hugging Face, along with implementation guidance and deployment options through Cohere’s Model Vault.

According to the company’s published specifications, the model uses a sparse MoE architecture, has 218 billion total parameters with 25 billion active parameters, supports a 128,000-token input context, and can generate up to 64,000 tokens. It is designed for text, image, and tool-use inputs and can produce text, reasoning, and tool-use outputs across 48 languages.

Command A+ Focuses On Enterprise Efficiency And Sovereign AI Deployment

Cohere said Command A+ is optimized for reasoning, retrieval-augmented generation, multilingual tasks, multimodal document processing, and agentic workflows. The company also emphasized its hardware efficiency, noting that the model is available in multiple quantization formats, including BF16, FP8, and W4A4, and can run on a small number of high-end GPUs. The model is also presented as Cohere’s fastest release to date, with the company citing improvements in output speed, lower latency, and better token efficiency compared with previous Command models. Cohere said speculative decoding has been adapted for the model’s architecture to further improve inference speed without affecting output quality.

The company described Command A+ as a consolidated successor to earlier Command A variants, stating that it surpasses previous generations on enterprise-focused tasks such as long-context reasoning, multimodal understanding, retrieval, and code generation. Cohere also highlighted benchmark gains over Command A Reasoning, including stronger results on telecom-related agentic evaluation, terminal-based coding tasks, and internal measures used within North. The model was also said to improve performance in spreadsheet analysis, agentic question answering, and memory-based workflows.

Multilingual capability is another central part of the release. Cohere said Command A+ expands support from 23 languages to 48, while improving translation and multilingual reasoning performance. The company also reported gains in non-European languages, including improvements in tokenization efficiency for Arabic, Korean, and Japanese. In multimodal benchmarks, the model was said to show stronger results on document and image reasoning tasks, including MathVista, MMMU, and CharXiv-related evaluations.

Industry interest in the release was reflected in comments from Fujitsu, which said the model’s architecture and agentic performance align with its own sovereign AI strategy. Cohere framed the launch as part of a broader effort to address three priorities in enterprise AI at the same time: capability, security, and cost. The company said future development will continue to focus on reasoning, multimodal performance, and coding while keeping models suitable for deployment inside customer-controlled environments.

Command A+ is available through Hugging Face, Model Vault, Cohere’s API, and a free online demo space. Cohere said documentation and deployment materials are also available for organizations looking to test or integrate the model into enterprise workflows.

The post Cohere Unveils Command A+, Advancing Sovereign And Efficient Enterprise AI Deployment appeared first on Metaverse Post.