AI company Cohere has introduced Command R+, a new large language model (LLM) designed for enterprise applications. The model specializes in Retrieval Augmented Generation (RAG) and tool usage.
The new model builds on the strengths of the recently introduced Command R and improves performance across the board. Cohere claims that Command R+ outperforms similar models in the same category and is competitive with more expensive models like GPT-4.
Key features of Command R+, which features a 128k-token context window, include:
- Advanced Retrieval Augmented Generation (RAG) with source citations to minimize misinformation
- Support for ten key business languages: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, and Chinese
- Integration with tools to automate complex business processes
According to Cohere, the RAG focus (database chat) makes the new model particularly attractive to companies because it improves response accuracy and provides source information to reduce hallucinations.
In a comparative test evaluating text flow, citation quality and overall utility, Command R+ outperformed competing models such as GPT-4. The model also demonstrated higher accuracy in question-and-answer tests using Wikipedia and Internet data, according to Cohere.
Cohere's pricing for Command R+ is competitive, with input and output costs for one million tokens set at $3 and $15, respectively. This is on par with Claude 3 Sonnet. The latest GPT-4 Turbo model costs $10 for one million input tokens and $30 for one million output tokens, Claude 3 Opus costs $15 and $75 respectively.
Language model with tool access
Command R+ also offers tool usage capabilities accessible through the Cohere and LangChain APIs. This allows the model to perform tasks such as automatically updating CRM records.
A new feature is the support for multistep tool use, enabling the model to combine several tools in multiple steps to solve complex tasks. Command R+ can even self-correct if it attempts to use a tool incorrectly or if the tool fails.
In tool benchmarks like Microsoft's ToolTalk (Hard) and Berkeley's Function Calling Leaderboard, Command R+ achieves results comparable to Claude 3 Sonnet and GPT-4 Turbo.
While Command R+ and its weights are freely available, the license is limited to research purposes only. Those interested in commercial use, either locally or on a cloud platform not currently offered, should contact Cohere directly for a license agreement. A demo version of the model is available for testing.
Cohere has also announced a collaboration with Microsoft Azure, enabling developers and enterprises to access its latest models through the Azure platform. In the coming weeks, Cohere will also be available on Oracle Cloud Infrastructure (OCI) and other cloud platforms.