The model infrastructure Cohere operates: what is deployed
Cohere is a Canadian company based in Toronto, and it occupies an unusual position in this reference: the only vendor that registers Persian item by item in a model language list, and the only vendor that has separately blocked all three routes to that Persian. One route blocked by pricing, one by licence, one by contract clause.
- Persian named in two models
- Three different licences
- Canadian infrastructure
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Model families
The Cohere model lines, from the commercial-tier Command to the research-tier Aya.
Versions
| Versions | Deployed | Status | Model families |
|---|---|---|---|
| Aya Expanse 32B | current | Aya | |
| Cohere Transcribe | current | Cohere Transcribe | |
| Cohere Transcribe Arabic | current | Cohere Transcribe | |
| Command A+ | current | Cohere Command | |
| Command A Reasoning | current | Cohere Command | |
| Command A Translate | current | Cohere Command | |
| Command A Vision | current | Cohere Command | |
| Command A | current | Cohere Command |
Persian is registered here by name, and that is rare
In this reference the usual answer to does this model process Persian is a guess. The vendor registers a round figure, a hundred languages say, and does not publish the list. Cohere has not followed that pattern in two places. Command A Translate is a dedicated machine translation model and it names its 23 languages one by one, with Persian the last name on that list. Aya Expanse 32B supports the same 23 languages and Persian is registered among them.
For an infrastructure that publishes in four languages, that list is more revealing than any benchmark: English, Arabic, Turkish and Persian, all four, inside one list of 23. This model was engineered for exactly the workload we run every day.
Three licences, three blocked routes
Now the second half. All three model units that touch Persian are deployed under a different licence, and each one fails for a different reason.
Command A Translate ships closed weights and Cohere registers no rate for it. Its pricing page registers a per-million-token rate only for units it deprecated on 15 September 2025, plus Aya Expanse. For Command A, Command A+, Translate, Reasoning and Vision no figure exists and the path routes you to sales.
Aya Expanse 32B publishes open weights, is 32 billion parameters, and is downloadable, but its licence is CC BY-NC 4.0 with an acceptable-use addendum. NC means no commercial use. Run a client site translation with it and charge for the work, and you are in breach. The same trap Voxtral TTS carries at Mistral is deployed here on a text model instead.
Cohere Transcribe is the only unit whose licence is both open and commercial: Apache 2.0, two billion parameters, speech recognition. And its list of 14 languages does not register Persian.
The contract registers Iran by name, and that is different
Every other vendor in this reference publishes a supported-countries list with Iran absent from it. Cohere took a different approach. Its commercial SaaS agreement defines a term, Restricted Location, and registers Iran in it alongside Belarus, China, North Korea, Russia and Syria, then states that the customer will not access the Services from a Restricted Location.
We log this separately because the weight of the claim differs. A name missing from a list can be an oversight. A name registered in a contract definition cannot be. So we log nothing about the Cohere hosted service and describe no route around it.
One detail we did not expect: Cohere is Canadian, its office is in Toronto and its agreement is governed by Ontario law, and it still carries this list. That shows how a sanctions list travels from one jurisdiction to another, and that an infrastructure being Canadian changes nothing on its own.
Where this infrastructure genuinely delivers: Arabic audio
This page is not only a list of blocked routes. One unit here is deployed nowhere else in this reference.
Cohere Transcribe Arabic is a dedicated finetune for Arabic audio, under Apache 2.0, with support for regional dialects and for switching between Arabic and English mid-sentence. No other vendor in this catalogue has deployed a speech recognition model built for Arabic. For a project with Arabic audio content, this is the best option in the reference, and its licence permits selling the result.
Two limits are self-documented and both matter: it produces no timestamps and does not separate speakers. For subtitling, that means manual processing.
What is not documentable in this architecture
The price of any current-generation Cohere unit. What the pricing page registers is the Model Vault rate, hourly rent on a dedicated instance: Embed 4 starts at $4.00 an hour and $2,500 a month, and Rerank 4 Pro at the large tier reaches $10.00 an hour and $6,500 a month. These are not token rates and do not compare with any other vendor, so the price column for Cohere models in our list stays empty.
And a small contradiction in their own figures: the Cohere research page registers that Aya Expanse covers 101 languages, while the API documentation for the same model registers 23. We register the documentation figure, because that is where the language list is also documented.
Documented strengths
- The only vendor in this reference to deploy a dedicated translation model that names its 23 languages one by one, Persian among them
- The only vendor in this reference to deploy a speech recognition model built for Arabic, under an Apache 2.0 licence at that
- It publishes the weights of its Aya research line together with parameter counts, a practice many vendors skip
- Its documentation registers a deprecation date for every retired unit, a practice few vendors follow
Known weaknesses and limits
- It registers no per-million-token rate for any current-generation unit. The only token rates on the pricing page belong to units deprecated on 15 September 2025.
- Its commercial SaaS agreement registers Iran by name in the Restricted Location definition, so we log nothing about the hosted service.
- Units whose weights are published openly are CC BY-NC, meaning commercial use is forbidden. Check this before building any product on one.
- Cohere Transcribe, the one with a commercial licence, does not support Persian, and it self-documents that it has no automatic language detection, so you must declare the language up front.
- The Arabic speech model produces no timestamps and does not separate speakers.
Technical verdict
If you have Arabic audio, Cohere is the best option in this reference and its licence permits selling the result. For every other workload, Cohere is an infrastructure worth reading, not one to deploy on: its language list is the most precise in this reference, and all three routes to it are blocked.
Frequently raised questions, with documented answers
Does Cohere operate from Iran
Cohere commercial SaaS agreement registers Iran by name in its Restricted Location definition and states that the customer will not access the Services from such a location. So we log nothing about the hosted service and describe no route around it. The Aya model weights are published separately, but their licence forbids commercial use.
Which Cohere model handles Persian
Two units. Command A Translate, a dedicated translation model that registers Persian in its list of 23 languages, and Aya Expanse 32B, which supports the same 23 and publishes its weights openly. The first has no rate, the second has no commercial licence.
Is it good for Persian speech to text
No. Cohere Transcribe names fourteen languages and Persian is not registered among them. For Arabic, however, it is the only dedicated unit in this reference and carries an Apache 2.0 licence.
Why is the price column empty for Cohere models in your list
Because Cohere publishes no token rate for any current-generation unit. Empty means there was no data to extract, not that it is free.