Cohere today announced the release of Command A+, an open-source large language model (LLM) built for high-performance agentic tasks. The model, which combines a mixture-of-experts (MoE) architecture, is available under an Apache 2.0 license and is designed for efficient deployment with minimal compute overhead. Command A+ surpasses previous Command models in key enterprise capabilities, including multimodal understanding, retrieval, and complex reasoning. The model supports 48 languages and is optimized for reasoning, agentic workflows, and multilingual document processing. Developers can access the model through Hugging Face, with weights available in several near lossless quantizations. For managed inference, Cohere recommends deploying Command A+ in Model Vault. Source: cohere
The model's architecture allows it to run on as little as two NVIDIA H100s or a single NVIDIA Blackwell GPU with minimal quality degradation. Command A+ is also the fastest model to date, achieving up to 63% higher Output Tokens per Second (TOPS) and reducing Time To First token (TTFT) by up to 17% compared to previous versions. The model's efficiency is further enhanced by speculative decoding, which provides an additional 1.5-1.6x inference speedup for both text and multimodal inputs. Source: cohere
Cohere emphasized that Command A+ represents a significant expansion of multilingual capabilities, increasing language support from 23 to 48 languages. The model also shows improvements in machine translation and multilingual reasoning. Tokenization efficiency improved by 20% for Arabic, 16% for Korean, and 18% for Japanese. Cohere noted that these gains extend to major non-European languages, which are often underrepresented in tokenizer training. Source: cohere
