Cohere has unveiled a major overhaul of its Command R and Command R+ AI models, significantly boosting their capabilities for enterprise applications. The latest iterations, command-r-08-2024 and command-r-plus-08-2024, mark a leap forward in retrieval-augmented generation, multilingual support, and adaptive tool use.
The standard Command R model now rivals its larger predecessor, Command R+, in overall performance. This upgrade brings a 50% increase in throughput and a 20% reduction in latency, all while halving its hardware footprint. These efficiency gains translate to faster, more cost-effective AI operations for businesses.
Supporting an impressive 23 languages, the new Command R series excels at multilingual RAG. This advancement is crucial for global enterprises, as it drastically reduces AI hallucinations by providing accurate, in-line citations across diverse languages. The models now seamlessly perform planning, tool querying, and question-answering in the user's preferred language without additional prompting.
Cohere has fine-tuned the models' capabilities in coding, mathematics, and logical reasoning. The improvements in structured data analysis are particularly noteworthy, enabling more sophisticated insights from various data formats, including tabular data. This enhancement is set to revolutionize data-driven decision-making across industries such as finance, SaaS, and consulting.
In a significant move towards customizable AI governance, Cohere has launched Safety Modes in beta. This feature allows enterprises to toggle between "Strict" and "Contextual" safety settings:
- Strict Mode: Enforces rigorous content guardrails, avoiding all sensitive topics. Ideal for corporate environments and customer service applications.
- Contextual Mode: Offers more flexibility for creative and academic uses, maintaining core protections while allowing a wider range of content.
This nuanced approach to AI safety empowers businesses to tailor model outputs to their specific risk tolerance and use cases, a crucial factor in building trust and ensuring responsible AI deployment.
The upgrade introduces new features to enhance usability and customization options. These include improved Structured Outputs for more accurate JSON generation and better instruction-following in preambles. The models also show reduced sensitivity to non-semantic prompt changes, offering more consistent results.
Cohere has made these powerful upgrades available at competitive rates:
command-r-08-2024: $0.15 per million input tokens, $0.60 per million output tokenscommand-r-plus-08-2024: $2.50 per million input tokens, $10.00 per million output tokens
The models are accessible via Cohere's hosted API and Amazon SageMaker, with plans to expand to additional cloud platforms.