Cohere, the Toronto AI company, sells a family of language models through an API and through major cloud platforms: Command models for text generation, Embed for search, Rerank for ordering results, Parse for documents, and Aya for multilingual research. The terms of use differ by model. Its May 2026 flagship, Command A+, is published with Apache 2.0 weights, while several earlier research models carry a non-commercial Creative Commons license.
Key takeaways
- Cohere's model catalog, as listed in its documentation on October 27, 2026, covers generation, embeddings, reranking, document parsing, speech transcription, and multilingual models.
- Command A+ is a mixture-of-experts model with 218 billion total and 25 billion active parameters, released as Apache 2.0 weights.
- Command A (March 2025) and Aya weights are listed on Hugging Face under CC BY-NC 4.0, which restricts commercial use.
- Models are offered on Cohere's own platform, Amazon SageMaker, Amazon Bedrock, Microsoft Azure, and Oracle's GenAI service.
- Older Command models were deprecated on September 15, 2025, so model names in tutorials can be out of date.
What the catalog contains
Cohere describes its models in a public model overview. The table groups the families and names one current example from each. Context lengths and names change often, so the documentation page is the authority.
| Family | Job | Current example (October 2026) |
|---|---|---|
| Command | Chat, tool use, agents, retrieval-augmented generation, translation | command-a-plus-05-2026, 128k context, text and image input |
| Embed | Turns text and images into vectors for search and classification | embed-v5.0-pro, 128k context |
| Rerank | Reorders documents returned by an existing search system | rerank-v5.0-pro, 32k context |
| Parse | Extracts structured data from forms and PDFs | parse-v5.0 |
| Transcribe | Speech to text, with an Arabic variant | cohere-transcribe-03-2026, 25 MB files |
| Aya | Wider language coverage | tiny-aya-global, 3.35 billion parameters, 70 languages |
| North | Purpose-built models | north-small-translate and north-mini-code |
Rerank and Embed are the less visible products. They sit inside search and retrieval systems that already exist, which is the usual way enterprises add a language model without replacing their stack.
Command A+ and its hardware footprint
The model card for Command A+ describes a mixture-of-experts design: 218 billion parameters in total, of which 25 billion are active for any one token. It accepts text and images, supports a 128k-token input, and is published in three precisions. The 4-bit version is the one the card recommends, and it lists a minimum of one NVIDIA B200 or two H100 GPUs for it. Cohere's documentation states that Command A, the March 2025 model, has a 256k context and needs two GPUs. These are vendor statements about its own models, and independent benchmark comparisons are outside the scope of this article.
Licenses differ from model to model
An "open weights" label hides real differences. The license field on each Hugging Face page, read on October 27, 2026, shows the split.
- Apache 2.0, which permits commercial use: Command A+ (all published precisions), the Cohere Transcribe models, and North Mini Code.
- CC BY-NC 4.0, which forbids commercial use: Command A, Command A Reasoning, Command A Vision, Command A Translate, the Aya Expanse and Aya Vision models, and the tiny Aya models.
A company that wants to run a CC BY-NC model on its own servers for a paid product would need a separate agreement or the hosted API instead. Anyone planning to build on a model should read the license text on the model's own page, because it can change between releases, as the move from Command A to Command A+ shows.
Where the models run and how access is priced
Cohere lists five ways to call its models: its own platform, Amazon SageMaker, Amazon Bedrock, Microsoft Azure, and Oracle's GenAI service. The pricing page separates trial keys, which are free but rate-limited and meant for prototypes, from production keys. Prices are billed per token and change, so none are quoted here.
The retirement schedule matters for anyone maintaining older code. According to the documentation, command-r-03-2024, command-r-plus-04-2024, the plain command and command-light models, and the generic aliases were deprecated on September 15, 2025, and the 8-billion-parameter Aya Expanse and Aya Vision models were retired on April 4, 2026.
Where the company fits
Cohere was founded in 2019 and is headquartered in Toronto. Its funding history, government compute support, and the announced combination with Aleph Alpha are covered in a separate article on this site. This page concentrates on the product line, since model names and licenses are what a developer or a procurement team actually has to check.
A short checklist before adopting any listed model
- Confirm the exact model name and whether it is still live or scheduled for deprecation.
- Read the license on the weights page and decide whether the intended use is commercial.
- Check whether the language of the content is on the model's supported list.
- Decide whether prompts may leave the organization, which decides between a hosted API and self-hosting.
- Test on a sample of real documents rather than relying on vendor descriptions.
FAQ
What does Cohere actually sell?
Access to its language models by API and through cloud marketplaces, plus the open-weight releases that anyone can download under the stated licenses. Its documentation lists generation, embedding, reranking, parsing, and transcription models.
Are Cohere models free to use commercially?
It depends on the model. Command A+ is released under Apache 2.0, while Command A and the Aya models are listed under a non-commercial Creative Commons license. The hosted API is billed separately.
What is the difference between Embed and Rerank?
Embed converts content into numerical vectors so similar items can be found. Rerank takes a list of results that a search system already returned and puts the most relevant first.
Can the models run on a company's own hardware?
The open-weight models can, subject to their licenses. The Command A+ card gives the example of one B200 or two H100 GPUs for its 4-bit version.
Related reading
- Cohere's 2023 funding round
- GPT-3 and early large language models
- Google's LaMDA dialogue model
- Characteristics of intelligent agents
More in AI Tools and Industry.
Last verified: October 2026
Sources
- An Overview of Cohere's Models, Cohere documentation, read October 27, 2026
- How Does Cohere's Pricing Work, Cohere documentation, read October 27, 2026
- Model Card for Command A+, CohereLabs on Hugging Face, May 2026
- Model Card for Command A (c4ai-command-a-03-2025), CohereLabs on Hugging Face, 2025
- License fields of CohereLabs model pages, Hugging Face, read October 27, 2026