A model designed for retrieval
On 11 March 2024, Cohere For AI — Cohere’s research arm — releases the weights of Command-R, a 35-billion-parameter model specifically designed for Retrieval-Augmented Generation and external tool use scenarios. 4 April 2024 follows with Command-R+, the flagship 104-billion-parameter version, with extended capabilities on the same use cases.
Unlike generalist models, Command-R is calibrated from training to handle long prompts containing retrieved documents, produce inline citations to sources and call external functions in a structured way. This optimisation is visible in RAG benchmarks and tool-based orchestration tasks, where the Command-R series shows superior performance compared to non-specialised models of comparable size.
Licence and distribution
The weights of Command-R and Command-R+ are distributed under the Creative Commons BY-NC 4.0 licence, which allows use for research and non-commercial purposes with an attribution requirement. For commercial use, Cohere requires API access or a dedicated contract. This choice differentiates the series from Apache 2.0 or MIT models, but still makes the full weights available for study, academic fine-tuning and non-commercial deployments.
Technical features
Both models offer a context window of 128,000 tokens, suitable for ingesting extensive documents without aggressive chunking. Multilingual support covers ten main languages — including English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic and Chinese — with quality intended for multi-region enterprise use cases.
The native citation mode allows the model, when it receives documents as input, to generate responses accompanied by explicit references to the relevant text portions, simplifying the verifiability and auditing requirements typical of knowledge management systems. Tool use is structured according to a scheme compatible with OpenAI function calling conventions, facilitating integration into existing stacks.
Link: cohere.com/command