Cohere: the Canadian maker of enterprise AI models
Cohere is a Toronto-based Canadian company that builds language models for businesses rather than consumers. It stands out for its search and retrieval models (Embed, Rerank) and for letting companies run its models in private environments.
Data as of October 22, 2026
Key figures
| Country | Canada |
|---|---|
| Headquarters | Toronto |
| Founded | 2019 |
| Type | Start-up |
| Weights policy | Mixed (open and closed) |
| Model families | Command, Aya, Embed, Rerank |
| Official site | cohere.com · Developer API |
| Models tracked here | 1 |
Roots in the Transformer
Cohere was founded in Toronto in 2019 by Aidan Gomez, Nick Frosst and Ivan Zhang. Aidan Gomez, the CEO, is one of the authors of the 2017 paper “Attention Is All You Need,” which introduced the Transformer architecture behind today’s language models. The company now has offices in New York, London, San Francisco, Montreal, Paris, Seoul and Berlin.
From the start, Cohere took a different path from OpenAI or Google: no consumer assistant, just models and tools for enterprises.
Model families built for document work
Cohere’s lineup reads like a complete processing pipeline:
- Command: generative models designed for agents, tool use and retrieval-augmented generation with source citations.
- Embed: models that turn text and images into embeddings for semantic search and classification.
- Rerank: reranking models that put the documents found by a search engine back in the right order.
- Aya: multilingual models from Cohere Labs’ open research.
On top of that come specialized models for document parsing (Parse), transcription (Transcribe) and translation. This combination explains Cohere’s place in internal search projects, where the right information has to be found before an answer is written.
Open weights for research, closed offering for business
Cohere’s approach is mixed. Cohere Labs, its open research lab launched in 2022, publishes open weights on Hugging Face, including for Command A. But those weights are under the CC-BY-NC license: non-commercial use only. For a paid product or service, you need a Cohere commercial plan.
How to access the models
- The Cohere API, using a key from the dashboard, documented at docs.cohere.com.
- Cloud marketplaces (AWS, Azure, Oracle, Google Cloud).
- Private deployment on dedicated infrastructure, in a VPC or on-premises.
- North, the enterprise platform launched in 2025, which adds ready-made agents.
To see how Cohere compares with other model makers, browse our AI labs page or the profile of Mistral AI, another non-US player focused on businesses.
Cohere models tracked on this site
| Models | Released | Capability index | Price (input / output) |
|---|---|---|---|
| North Small Translate | September 9, 2026 | — | — |
Frequently asked questions
Does Cohere offer a chatbot like ChatGPT?
Not for consumers. Cohere sells North, an AI platform for workplace use, plus API access to its models. Developers also get a playground to test them.
Are Cohere models open source?
Partly. Cohere Labs, the research arm, publishes model weights such as Command A and Aya on Hugging Face, but under the CC-BY-NC license, which forbids commercial use. For business use you need Cohere's commercial offering.
Can I use Cohere without sending data to the public cloud?
Yes, that is one of its main selling points. Cohere offers deployments on dedicated infrastructure (Model Vault), in a virtual private cloud (VPC) or on your own servers.
Where can I find Cohere models besides its own API?
According to its documentation, on several cloud marketplaces: Amazon Bedrock and SageMaker, Microsoft Azure, Oracle OCI and Google Cloud. The models on offer vary by platform.
Data sources
- cohere.com
- cohere.com
- docs.cohere.com
- docs.cohere.com
- huggingface.co
- Models.dev — Model data: models.dev (open-source project)
- Epoch AI — AI Benchmarking Hub — Epoch AI, “AI Benchmarking Hub”, epoch.ai, CC BY 4.0
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.