What is Command A+?

Command A+ is Cohere’s fastest and most powerful language model yet—an open-source enterprise workhorse built for complex reasoning, multimodal and multilingual agentic tasks. Released under an Apache 2.0 license, this mixture-of-experts (MoE) model delivers 218B total parameters with only 25B active, enabling high-performance agentic workflows on as little as two H100 GPUs. It unifies capabilities from across the Command family into a single scalable model, supporting text, image, and tool use inputs with a 128K input context and 64K maximum generation length.

Who it's for

• Enterprise developers who need a privately deployable LLM for building sovereign agentic AI systems without relying on external APIs.
• AI researchers and engineers experimenting with open-source models, requiring efficient inference, low-bit quantization, and integration with frameworks like vLLM and Transformers.
• Multilingual teams handling complex reasoning, retrieval-augmented generation (RAG), and multimodal document processing across 48 supported languages.

Key features

Mixture-of-Experts architecture for efficiency

Command A+ uses a sparse MoE design with 218B total parameters but only 25B active per forward pass. This dramatically reduces compute requirements while maintaining enterprise-grade performance, making it deployable on minimal hardware like two H100 GPUs.

Unified reasoning, multimodal, and tool use

Unlike previous Command A models that specialized in separate capabilities, Command A+ combines reasoning, vision understanding, tool use, and multilingual support in one model. It handles complex agentic tasks, long-horizon planning, and multimodal document processing without needing separate models for each function.

Optimized for sovereign deployment

The model is freely available under Apache 2.0 with support for near-lossless quantizations and integration with open inference frameworks. Developers can download weights from Hugging Face and deploy in their own environments, including Cohere’s Model Vault for managed inference.

What stands out

Command A+ makes sovereign agentic capabilities available to all—running on as little as two H100 GPUs while outperforming every previous generation in the Command series.

This combination of open-source accessibility and enterprise-grade performance is rare. The model achieves dramatic improvements over its predecessors, with agentic coding performance on Terminal-Bench Hard jumping from 3% to 25% and telecom reasoning scores rising from 37% to 85%. It’s designed from real-world deployment experience with North, Cohere’s enterprise workspace, giving it practical grounding in production agentic workflows.

Worth checking out if…

You’re building agentic AI systems that need to run privately, efficiently, and across multiple languages. Command A+ is ideal if you want a single model that handles reasoning, vision, tool use, and multilingual tasks without juggling separate specialized models. It’s particularly valuable for enterprises prioritizing sovereign AI deployment with minimal compute overhead.