Cohere Open-Sources Command A+: 218B Parameter MoE Model Focused on Enterprise Agents and Data Sovereignty

According to monitoring by Beating, Cohere has officially open-sourced the sparse mixture-of-experts model Command A+ with 218 billion (218B) parameters under the Apache 2.0 license, supporting enterprise-level agent development and private deployment. The new model is specifically designed to meet the data sovereignty needs of critical infrastructure, enabling physical isolation deployments without vendor lock-in. While the total parameter count is 218 billion (218B), only 25 billion (25B) parameters are activated per inference, achieving a balance between computational efficiency and generative performance. The optimized inference efficiency allows it to run on just two NVIDIA H100 GPUs or a single B200 GPU. Hugging Face has also launched low-precision quantized versions such as W4A4, further lowering the deployment threshold. Command A+ natively supports multimodal image-text input, offering a 128K-token input context window and a 64K-token output length. It has been deeply optimized for complex logical reasoning, autonomous tool usage, database queries, and other agentic workflows, as well as long-document processing. Additionally, it supports 48 languages, including all official EU languages. [Original Link]