Bottom line

Cohere's Command A+ enters the enterprise fray as a multimodal, open-weight (Apache 2.0) Sparse Mixture-of-Experts model. By combining a 128K input context, tool use, and native citation grounding with a hardware-efficient footprint, it targets the sweet spot of enterprise sovereignty and frontier-level utility.

Signal

The clearest signal here is Cohere's commitment to the private, sovereign AI lane. Releasing an Apache 2.0 licensed frontier model with 218B total parameters—but only 25B active during inference—means operators get top-tier reasoning without needing a hyperscaler-sized GPU cluster. Confirmed to fit on one B200 or two H100 GPUs, it is explicitly designed for private and on-premise use. It natively supports both text and vision inputs while generating text, providing built-in grounded outputs and tool use for complex workflows.

Noise

While the active parameter count makes it highly efficient, labeling it strictly by its 218B total parameter footprint might confuse operators regarding its actual compute requirements. The noise is in the raw parameter count versus the operational reality: it is a highly optimized MoE built specifically to fit into constrained enterprise hardware bounds.

Model profile

• Provider: Cohere
• Architecture: Sparse Mixture-of-Experts (218B total / 25B active parameters)
• Modality: Multimodal (text and vision inputs, text outputs)
• Context Window: 128K input, 64K output
• License: Open-weight, Apache 2.0
• Release Date: May 20, 2026

Assessment

Command A+ is designed to solve a specific enterprise headache: the trade-off between model capability and data privacy. By releasing under an Apache 2.0 license, Cohere allows businesses to take this model completely behind their own firewalls. The native 48-language support and deep integration with Cohere's chat and agents stack demonstrate that this isn't just a raw foundational model; it is packaged for immediate, global agentic application.

Where it fits

This model is built for environments where data cannot leave the building. It fits perfectly into on-premise enterprise deployments requiring complex reasoning and tool use. With its 128K input window and a notably large 64K maximum generation limit, it is uniquely suited for digesting massive corporate documents and generating extensive, long-form reports backed by native citation grounding. The vision-language capabilities also position it well for multimodal document ingestion tasks across global business units.

Operator implications

The operator read is that Cohere is shifting the enterprise battleground from managed API endpoints to sovereign deployment. If your organization has strict compliance, data residency, or privacy mandates, Command A+ offers a pathway to frontier-like capabilities without sending data to an external provider. Furthermore, the massive 64K output generation is a strong directional signal: the emerging pattern is that enterprise models are increasingly expected to generate complex, multi-page artifacts rather than just brief conversational turns.