Cohere Releases Command A+ Open-Source: MoE Model for Enterprise Agentic AI

Summary

Cohere has released Command A+ as an open-source model under an Apache 2.0 license, marking a significant step toward sovereign AI. Command A+ is a mixture-of-experts (MoE) model designed specifically for high-performance agentic tasks with minimal compute overhead, built from a year of deploying the North enterprise workspace.

The model consolidates capabilities from across Cohere's Command family into a single, scalable solution. Command A+ significantly outperforms previous Command models across key dimensions, including agentic coding (improving from 3% to 25% on Terminal-Bench Hard), memory reasoning (from 39% to 54%), and spreadsheet analysis (by 32%). The model also substantially expands multilingual capabilities from 23 to 48 languages while achieving strong performance on multimodal understanding benchmarks.

Optimized for practical developer use, Command A+ supports low-bit quantization, efficient inference, and integration with open inference frameworks. The model weights are available on Hugging Face in multiple quantizations, with a managed inference option through Model Vault. Cohere frames the release as essential infrastructure for AI independence, enabling developers to run, control, and adapt models without vendor lock-in.

• Supports low-bit quantization and efficient inference for local deployment; weights available on Hugging Face and via Model Vault

Editorial Opinion

Command A+ represents Cohere's clearest commitment yet to sovereign AI as a platform differentiator. By releasing a consolidated, openly deployable model trained on enterprise agentic workloads, Cohere is directly addressing the most immediate concern holding back enterprise AI adoption: vendor lock-in and the inability to run models independently. The performance gains across agentic tasks and the expansion to 48 languages signal that Cohere is building for global enterprise deployment, not research benchmarks.