Cohere’s model licences fall into three groups: proprietary models used only through Cohere’s hosted service, models whose weights are published under a permissive Apache 2.0 licence, and models whose weights are published under the Creative Commons Attribution-NonCommercial 4.0 licence. Which group a model belongs to decides whether an organization may download the weights, whether it may use them commercially and which agreement governs the use. This article records what Cohere’s own documents say about each.
Editions
The pricing page shows a licence tag on each model card. Command A+ is tagged Apache 2.0, one reranker model card is tagged Proprietary, and the Aya Expanse 32B, Aya Vision and Tiny Aya cards are tagged CC-BY-NC-4.0; the tags load as page data rather than visible text, so the model documentation pages below are the readable confirmation.[1]
Apache 2.0 releases
- Command A+ is Cohere’s first mixture-of-experts model and its model card states that it is available under an Apache 2.0 licence on Hugging Face.[3] Cohere says it is free through the API until rate limits are reached and that production use is through Model Vault.[3]
- North Mini Code is a 30 billion parameter model with 3 billion active parameters for agentic coding, released under the Apache 2.0 licence. It is free through the API until rate limits are reached and supports production deployment through Model Vault.[4]
- Cohere Transcribe is described as an open source research release of a 2 billion parameter speech recognition model under Apache 2.0, available through the API at no cost for experimentation subject to rate limits, with Model Vault for production deployment without rate limits.[6]
Apache 2.0 allows commercial use, but the licence travels with the downloaded files. The licence text is not hosted on the Cohere pages cited here, so an asset manager should capture the LICENSE file shipped with the specific weights and record it with the model version, rather than relying on this summary.
CC BY-NC 4.0 releases
- North Small Translate is a mixture-of-experts machine translation model with 218 billion total and 25 billion active parameters. Cohere states that its weights are openly available in W4A16, FP8 and BF16 formats under the Creative Commons Attribution-NonCommercial 4.0 licence for non-commercial use.[5] Cohere offers a free tier on the API, subject to the limitations in the applicable commercial agreements, including the Usage Policy, and directs production use to Cohere sales.[5]
- Aya Expanse 32B and Aya Vision. Cohere Labs released Aya Expanse 32B and Aya Vision as open-weight models through Hugging Face for research purposes.[11][14] The pricing page tags the Aya Expanse 32B, Aya Vision and Tiny Aya cards CC-BY-NC-4.0.[1]
- Tiny Aya is a 3.35 billion parameter multilingual model released as a family of five models: a pretrained base model, a globally balanced instruction-tuned variant and three region-specialized variants.[10]
The non-commercial restriction is a property of the weights. An organization that wants commercial use of the same capability has two routes in Cohere’s documents: the API, where Aya Expanse models are charged at USD 0.50 per million input tokens and USD 1.50 per million output tokens,[1] or a commercial arrangement with Cohere sales for models such as North Small Translate.[5] A risk specific to this group is shadow use: a research team downloads weights for a pilot and the pilot becomes a production workload. The catalog row is Open-weight releases.
Proprietary models
Models such as Command A, Command R, Command R7B, Command A Reasoning and the Embed, Rerank and Parse families are consumed through Cohere’s API, Model Vault, cloud marketplaces or a private deployment licence rather than as downloadable open weights.[2] Their use is governed by the agreement for the route; see Cohere agreements and data terms.
Metrics
Open-weight models carry no per-use fee from Cohere when self-hosted; the cost is the customer’s own compute. When the same model is used through Cohere, the metric is the one for that route: billed tokens on the API, or the instance on Model Vault, where each model is listed per instance-hour by performance tier; see Cohere Model Vault and private deployments. The relevant rows are Chat tokens and Model Vault instance.
Counting and floors
Cohere’s open-weight pages do not state seat, device or instance limits on downloaded weights; whatever Apache 2.0 or CC BY-NC 4.0 provides is the full set of use rights. The Client License Agreement, by contrast, restricts commercial products obtained through a partner: it prohibits making them available to third parties, and North used with a model other than Cohere’s removes Cohere’s indemnity.[13]
Policies that follow the model
Usage Policy
The Usage Policy applies to all Cohere products and services, including models, software, applications and the API.[8] Its universal requirements prohibit, among other things, child sexual exploitation, incitement of violence, illegal activities including generating outputs that may infringe third-party intellectual property, compromising privacy or security, surveillance and predictive policing, deceptive practices, misinformation and abusing Cohere services through jailbreaking or prompt injection.[8] It also prohibits high-risk activities, including automated determinations about individuals in domains such as employment, education, healthcare, housing, law enforcement and financial or insurance services, while stating that back-office uses such as document summarisation, transcription and internal knowledge agents are not high risk.[8]
A customer that builds an application on Cohere must ensure it complies with the universal requirements and that its own users are bound to substantially similar requirements. Public-facing applications that interact with people, such as chatbots and agents, must disclose that the user is dealing with an AI system.[8] The SaaS Agreement and the Client License Agreement both incorporate the Usage Policy as a condition of use.[12][13] The catalog row is Usage Policy.
Cohere Labs Acceptable Use Policy
Cohere Labs publishes a separate acceptable-use policy for its models and model derivatives. It expects users to comply with applicable laws and prohibits violence and harm, harassment and abuse, sexual exploitation, collecting or inferring sensitive personal information without lawful authority, fraud and deception, and generating synthetic data outputs for commercial purposes, including to train, improve, benchmark or otherwise develop model derivatives or related products or services.[7] The synthetic-data prohibition is the clause most likely to affect an engineering team: using a Labs model to produce training data for a commercial model is outside the policy. The catalog row is Cohere Labs Acceptable Use Policy.
Outputs and training data
The SaaS Agreement prohibits using the services or any Output to build or train a similar or competitive product or service, such as an AI model competitive with Cohere models, and treats any entity that offers large language models for licence or sale as a direct competitor.[12] That restriction applies to the hosted service, not to Apache 2.0 weights. Cohere also states that its models are trained on a proprietary mix of public information, datasets developed by Cohere with human annotation or synthetic generation, and datasets from specialised data vendors, and that its models are not for direct use by consumers or for personal or household purposes.[9]
Programs
Cohere’s research arm Cohere Labs releases open-weight models alongside the commercial range. The practical programme for a licence manager is an intake check: when a team asks to download Cohere weights, record the model name and version, the licence on the model card and in the downloaded files, and whether the intended use is commercial. Weights under CC BY-NC 4.0 are for non-commercial use only, and anything else routes to the API, Model Vault or a sales agreement.
Out of scope
The full text of the Apache 2.0 and Creative Commons licences, licences of fine-tuned derivatives made by third parties, and weights hosted on third-party sites, which this article does not cite because those sites are not Cohere documents. The related legal dispute over training data used by Cohere is covered in Advance Local Media v. Cohere.