Introducing Command A+ | Cohere

What this post added

This post introduces Command A+, a new Mixture-of-Experts (MoE) model that consolidates and enhances capabilities from previous Command generations. Key technical contributions include its sparse MoE architecture (218B total, 25B active parameters), improved performance on agentic tasks (e.g., 𝜏²-Bench Telecom, Terminal-Bench Hard, North internal evaluations), multimodal understanding (MMMU Pro, MMMU, MathVista, CharXiv), and multilingual support (48 languages). The post details significant engineering efforts in hardware efficiency through 16-bit, 8-bit, and 4-bit quantizations, leading to higher Output Tokens per Second (TOPS) and reduced Time To First Token (TTFT). It also highlights the use of speculative decoding optimized for MoE architecture and a new tokenizer that improves compression and tokenization efficiency for various languages. The model is designed for private deployment and integrates with open inference frameworks.