Cohere Labs Opens Command-R to Build Grounded Enterprise AI Agents
Cohere Labs releases Command-R v01, a 35B open-weights model purpose-built for RAG pipelines, tool use, and multilingual enterprise workflows
Takeaways
- Open weights release: Cohere Labs publishes Command-R v01, a 35B model purpose-built for RAG and tool use.
- 128K context window: Supports long-document grounding with inline citations baked into the training objective.
- Dual tool use modes: Supports single-step function calling and multi-step agentic loops (Action → Observation → Reflection).
- Multilingual by design: Optimized for 10 languages including Arabic, Japanese, Korean, and Chinese; pretrained on 23+ languages.
- CC-BY-NC license: Free for research; commercial use requires a separate agreement with Cohere.
- Multiple serving options: Works with Transformers, vLLM, SGLang, and Docker; a 4-bit quantized version is also available.
Cohere Labs Command-R (c4ai-command-r-v01) is now publicly available on Hugging Face. It is a 35-billion parameter open-weights model built specifically around the workflows that matter most to enterprise teams: retrieval-augmented generation (RAG), function calling, and multi-step agentic tool use. This is not a general-purpose chat model trying to do everything. It is a focused, production-oriented release from Cohere's research arm.
The problem it was built to solve
Command-R is a 35B model designed for production-scale AI in enterprises. It is a scalable generative model that balances efficiency with accuracy, enabling companies to move beyond proof-of-concept and into production. The core insight behind the design is that most enterprise AI failures happen not at the chat layer but at the retrieval and grounding layer: models hallucinate, lose track of sources, or fail to call the right external tool at the right time. Command-R addresses all three.
What it actually does well
Command-R boasts high precision on retrieval augmented generation (RAG) and tool use tasks, low latency and high throughput, a long 128,000-token context length, and strong capabilities across 10 key languages. The multilingual coverage is not an afterthought either.
The 10 primary languages optimized for performance are:
- English, French, Spanish, Italian, German
- Brazilian Portuguese, Japanese, Korean
- Simplified Chinese, Arabic
Command-R is a large language model with open weights optimized for a variety of use cases including reasoning, summarization, and question answering. Command-R has the capability for multilingual generation evaluated in 10 languages and highly performant RAG capabilities.
Grounded generation: citations baked into the model
Command-R has been specifically trained with grounded generation capabilities. This means that it can generate responses based on a list of supplied document snippets, and it will include grounding spans (citations) in its response indicating the source of the information. This behavior has been trained into the model via a mixture of supervised fine-tuning and preference fine-tuning, using a specific prompt template.