Overview
Command R (03-2024) is the original release in Cohere's Command R line, launched on 11 March 2024 by CEO Aidan Gomez as a model built for enterprise retrieval-augmented generation (RAG) at production scale. It is a 35-billion-parameter model with a 128K-token context window, served on the Cohere API as command-r-03-2024 (alias command-r). Cohere positioned it in the emerging 'scalable' tier that trades a little raw capability for much better efficiency and price, so companies could move RAG and tool-use applications from proof of concept into production.
Command R (03-2024) was designed around two enterprise workloads: grounded RAG, where it pairs with Cohere's Embed and Rerank models and returns answers with inline citations back to source passages to reduce hallucination, and conversational tool use, where it can call external tools, APIs, search engines, and databases, including chaining tools over multiple steps. It delivers strong performance across 10 key languages (English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, and Simplified Chinese), with additional languages seen during pre-training.
Cohere's research lab released open weights for this version on Hugging Face (CohereLabs/c4ai-command-r-v01) under a non-commercial CC-BY-NC license, while commercial deployments ran through the paid API. The model is text-only for both input and output, with on-demand responses capped at 4K output tokens. Command R (03-2024) was superseded by the August 2024 refresh and the larger Command R+, and Cohere deprecated it on 15 September 2025, pointing users to command-r-08-2024, command-r-plus-08-2024, or Command A.
| Released | 2024-03-11 |
|---|---|
| License | Open research weights on Hugging Face under CC-BY-NC 4.0 plus Cohere's Acceptable Use Policy (non-commercial); commercial use was offered through the Cohere API under separate commercial terms. |
| Weights | Open weights |
| Parameters | 35B |
| Context | 128K |
| Max output | 4K |
| Architecture | Auto-regressive language model using an optimized transformer architecture, post-trained with supervised fine-tuning (SFT) and preference training to align behavior for helpfulness and safety. It was purpose-built for retrieval-augmented generation with inline citations and for conversational, multi-step tool use (function calling). |
| Knowledge cutoff | 2024-03 |
| Modalities | Text |
| Status | Deprecated/retired. Cohere announced the deprecation of command-r-03-2024 (and the command-r alias) effective September 15, 2025, recommending command-r-08-2024, command-r-plus-08-2024, or command-a-03-2025 as replacements. Fine-tuned models built on command-r-03-2024 were sunset earlier, on March 8, 2025. |
Benchmarks
- MMLU-Pro37.9%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.50 / 1M tokens per 1M tokens |
|---|---|
| Output | $1.50 / 1M tokens per 1M tokens |
Cohere API on-demand pricing for the original command-r (03-2024), as aggregated by Artificial Analysis across providers. On-demand responses were capped at 4K output tokens. The model was deprecated on 15 September 2025.
Strengths
- Retrieval-augmented generation with grounded answers and inline citations to source passages
- Conversational tool use, including multi-step tool use that chains several tools to complete a task
- Large 128K-token context window for long documents and extended chat history
- Multilingual coverage across 10 key languages
- Open research weights on Hugging Face for study, fine-tuning, and self-hosting
- Low API pricing ($0.50 / $1.50 per million tokens) aimed at high-volume production RAG
- Pairs natively with Cohere's Embed and Rerank models for an end-to-end RAG stack
Best for
- Retrieval-augmented generation over enterprise knowledge bases and document collections
- Agents and assistants that call tools, APIs, search, and databases (single- and multi-step)
- Long-document question answering and summarization using the 128K context window
- Multilingual customer support and question answering across 10 languages
- Grounded, citation-backed answers where source attribution matters
- Cost-sensitive, high-throughput production pipelines where a larger model is overkill
How to access
| Provider | Model ID |
|---|---|
| Cohere ↗ | command-r-03-2024 |
Command R — every version
The full lineage of the Command R line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Command R (08-2024)current | 2024-08-30 | — | Proprietary |
| Command R (03-2024) | 2024-03-11 | — | Proprietary |
FAQ
What is Command R (03-2024)?
It is the original model in Cohere's Command R line, released on 11 March 2024. It is a 35-billion-parameter large language model with a 128K-token context window, built for enterprise retrieval-augmented generation (RAG) with citations and for conversational, multi-step tool use. On the Cohere API it was served as command-r-03-2024.
Is Command R (03-2024) still available?
No. Cohere deprecated command-r-03-2024 (and the command-r alias) effective 15 September 2025, recommending command-r-08-2024, command-r-plus-08-2024, or command-a-03-2025 instead. Models fine-tuned on command-r-03-2024 were sunset earlier, on 8 March 2025.
Were the weights open?
Yes, for research. Cohere's research lab published the weights on Hugging Face as c4ai-command-r-v01 under a non-commercial CC-BY-NC license plus its Acceptable Use Policy. Commercial use was offered separately through the paid Cohere API.
How much did Command R (03-2024) cost on the API?
Aggregated pricing put it at about $0.50 per million input tokens and $1.50 per million output tokens, with on-demand responses capped at 4K output tokens. It was the cheaper, smaller counterpart to the larger Command R+.