Command R+

Cohere built this 104-billion-parameter open-weights model for grounded answers that cite their sources and for agent loops that call several tools in sequence. Give it a conversation plus short document snippets and it works out which ones matter, writes a reply, and marks the passage behind each claim; a separate prompt format covers function calling and multi-step planning. The context window reaches 128K tokens, and Cohere tuned it for ten languages, among them Japanese, Korean, Arabic and Brazilian Portuguese. Weights download from Hugging Face after a short gated form and load through the transformers library, community GGUF quantizations exist for llama.cpp, and an August 2024 refresh sits beside the original. In 16-bit form the files total around 200 GB, so local runs mean a multi-GPU server or heavy 4-bit quantization on a high-memory workstation. The license is Creative Commons BY-NC 4.0 plus Cohere's acceptable use policy: research and personal projects are fine, commercial products must use the paid Cohere API. Best suited to RAG researchers and teams comparing citation behavior across open models.