Verified Performance Data

Cohere

The Command model family: From Command R to Command A. Enterprise-focused AI with strong retrieval-augmented generation. Benchmark performance across knowledge, reasoning, coding, math, multimodal, and instruction following.

3
Models Tracked
2019
Founded
88.2
Best Knowledge
90.9
Best Instruction

About Cohere

Cohere is a Canadian AI company founded in 2019 by former Google Brain researchers, including Aidan Gomez—one of the co-authors of the original “Attention Is All You Need” Transformer paper. Based in Toronto, Cohere focuses on enterprise AI with a strong emphasis on retrieval-augmented generation (RAG) and multilingual capabilities.

The Command model family is designed specifically for enterprise use cases: tool use, document processing, and multi-step agentic workflows. Command A (2025) represents Cohere’s latest generation, offering GPT-4-tier performance at significantly lower computational cost with a 111B parameter MoE architecture.

Command Model Timeline

The evolution of Cohere’s Command family. Verified scores from the Command A Technical Report.

March 2025

Command A

Cohere’s most capable model. Command A achieves 85.5% on MMLU, 50.8% on GPQA, and 90.9% on IFEval. A 111B parameter MoE architecture delivers strong performance with efficient inference, purpose-built for enterprise RAG and agentic workflows.

MMLU 85.5 GPQA 50.8 IFEval 90.9
April 2024

Command R+

The first model to bring Cohere into frontier territory. Command R+ achieved 88.2% on MMLU and introduced advanced multi-step tool use and strong RAG grounding, establishing Cohere as a serious enterprise AI contender.

MMLU 88.2
March 2024

Command R

An efficiency-focused model designed for production RAG workflows. Command R offered strong retrieval-augmented generation at lower latency and cost, making it popular for enterprise search and document analysis applications.

Benchmark Performance

Command A scores across verified benchmark categories.

Key Strengths

Instruction Following Leader

Command A achieves 90.9% on IFEval, one of the highest instruction-following scores across all providers—critical for enterprise applications where precise adherence to complex instructions matters.

Enterprise RAG Focus

Cohere models are purpose-built for retrieval-augmented generation, with native grounding, citation generation, and multi-step tool use designed for real-world enterprise document workflows.

Transformer Origins

Co-founded by Aidan Gomez, co-author of “Attention Is All You Need,” Cohere brings deep architectural expertise to model design, with a focus on practical deployment efficiency.

About This Data

Benchmark scores are sourced from Cohere’s official Command A Technical Report. Command R+ MMLU score is from Cohere’s model card. Note that Command A reports GPQA (not specifically GPQA Diamond); scores may not be directly comparable to providers that report the Diamond subset.

Explore More Providers

Compare Cohere’s Command models against other frontier AI systems.