Introducing Cohere Command-R-Plus-08-2024: Advanced LLM with Enhanced Capabilities

The latest update in Cohere's generative AI model series, CohereForAI/c4ai-command-r-plus-08-2024, also known as Command R+, brings significant improvements and new features designed to enhance your AI-driven tasks. Here's what you need to know:

Model Overview

Command R+ boasts a model size of 104 billion parameters, trained on 23 languages and evaluated in 10. It's optimized for reasoning, summarization, and question answering, supporting Retrieval Augmented Generation (RAG) and tool use for automating sophisticated tasks.

Technical Enhancements

• Performance: Delivers 50% higher throughput and 25% lower latencies compared to its predecessor, maintaining the same hardware footprint.
• Tool Use: Improved decision-making around tool use, supporting multi-step processes and including a directly_answer tool for straightforward questions.
• Multilingual RAG: Enhanced multilingual RAG searches, providing better responses in the user's language.
• Structured Data: Better analysis and creation of structured data from unstructured natural language instructions.
• Robustness: Improved robustness to non-semantic prompt changes, better handling of unanswerable questions, and improved citation quality.

Safety Modes

Command R+ introduces Safety Modes to provide context-sensitive safety guardrails:

• STRICT: Avoids all sensitive topics, ideal for general and enterprise use.
• CONTEXTUAL: Default mode, suitable for wide-ranging interactions with fewer constraints while maintaining core protections.
• NONE: Option to opt out of safety modes beta.

Usage and Integration

The model uses a specific prompt template for tool use, and deviating from this template may reduce performance. It integrates seamlessly with Hugging Face's transformers library (version 4.39.1 or higher). Example code is provided for rendering tool use prompts and basic usage.

Pricing

Command R+ is competitively priced at $2.50 per million input tokens and $10.00 per million output tokens.

Additional Features

• Safety Protections: Built-in protections against content endangering child safety, which cannot be adjusted.
• License: CC-BY-NC, adhering to C4AI's Acceptable Use Policy.

This model is part of Cohere's open weights research release and is designed to enhance various AI tasks with advanced tool use and safety features.