Cohere Releases Open-Source AI for Sovereign Cloud

Cohere has released Command A+, a groundbreaking 218 billion parameter AI model optimized for enterprise needs. This model is notable for its architecture, utilizing a sparse Mixture-of-Experts (MoE) design where only 25 billion parameters are active per generation step. This efficiency is further enhanced by advanced 4-bit quantization, allowing the model to run on minimal hardware, such as a single NVIDIA Blackwell B200 GPU.

Command A+ is released under the permissive Apache 2.0 open-source license, a significant move by Cohere. This licensing ensures enterprises, governments, and developers can operate, control, and adapt advanced AI within their own secure environments without sacrificing performance, embodying the concept of 'sovereign AI.' This offers total vendor independence and bypasses traditional API access limitations.

The model excels in agentic workflows, complex reasoning, and multimodal document processing. It demonstrates significant benchmark improvements in areas like mathematics and telecom reasoning. A key feature is its native citation generation, providing explicit grounding spans for factual claims, which is crucial for regulated industries.

Furthermore, Command A+ boasts native support for 48 languages with its overhauled tokenizer, significantly improving efficiency for non-European languages, thus reducing operational costs for global deployments. Its multimodal capabilities allow native processing of text and images within a massive 128K input context window.