Cohere Open-Sources Command A+, a 218B Moe Model That Runs on Two H100s (firethering.com)

Cohere has announced the open-source release of Command A+, a 218 billion parameter mixture-of-experts (MoE) model designed for enterprise tasks. This single model consolidates five previously separate models from the Command A family into one, which significantly improves efficiency and performance across various workflows. Command A+ boasts a 20% boost in agentic question-answering accuracy, a 32% improvement in spreadsheet analysis quality, and enhanced memory performance, with its ability to reference context from earlier sessions showing a jump from 39% to 54%. Notably, this model only needs two NVIDIA H100s for operation, allowing organizations to manage sensitive data on their own infrastructure without relying on external APIs. The introduction of Command A+ signifies a major advancement for the AI/ML community, especially for enterprise applications that require multilingual and multimodal capabilities. With support for 48 languages and improvements in reasoning for non-European languages, the model targets global deployments while streamlining the infrastructure requirements for businesses. While Command A+ excels in agentic tasks, its design focuses on specific enterprise functions rather than general chat capabilities. Open-sourcing the model under Apache 2.0 allows enhanced visibility into its architecture, fostering potential community contributions and driving innovation in the sector.