Cohere Releases Command A+, an Open Source Multimodal Reasoning Model

Cohere has announced in a blog post the release of Command A+, an open source language model built for multimodal and multilingual agentic tasks. The model is available under the Apache 2.0 license and can be deployed locally or through Cohere's Model Vault service.

Command A+ uses a mixture of experts architecture with 218 billion total parameters, of which 25 billion are active. It supports text, image, and tool-based inputs and outputs, and operates efficiently on as little as two NVIDIA H100 GPUs or one Blackwell GPU. The model is optimized for reasoning, retrieval, and document processing across 48 languages.

According to Cohere, Command A+ outperforms previous Command models in reasoning, multimodal understanding, and coding benchmarks. It achieves up to 63 percent higher token throughput and 17 percent lower latency compared with Command A Reasoning. The model introduces a new tokenizer that improves efficiency, reducing token usage by up to 20 percent for some languages.

Command A+ is available for download on Hugging Face in several quantizations, including 16-bit, 8-bit, and 4-bit formats. Developers can also access implementation guides and deployment documentation through Cohere’s platform.