Enterprise
For teams serving Tunisia.
Your customers write and speak in Derja. We build the models for it, we are in Tunis, and the contract, the invoice and the person who answers are all here.
Who this is for
Institutions whose customers do not write Fusha.
Most Arabic language models are trained on Modern Standard Arabic. That is the language of the news, of textbooks and of official letters. It is not the language anyone banks in, complains in, or leaves a voice note in.
A contact center
Calls switch between Derja and French. Start with transcription and agent-reviewed summaries, then measure the corrections and time saved. Read the pilot guide.
A bank
A customer writes to support in Derja with three French words in the middle of the sentence. That message has to be read, routed and answered, and the transcript has to be readable by whoever audits it later.
A telco
Thousands of calls a day, most of them short, almost none of them in a language a general speech model was trained to hear. The cost of getting them wrong is a person repeating themselves to a machine.
A public service
Citizens who do not write French and should not have to. Here the terms matter as much as the model: what is kept, for how long, and who can be shown the answer.
What there is to discuss
None of this is bought off a page.
Every item below is different for a bank than for a telco, which is why there is no form to fill in and no tier to pick. This is the agenda for a first conversation.
Where it runs
Shared capacity is what the public API uses today. Dedicated capacity and running inside your own infrastructure are both real shapes for this and both are conversations rather than switches. What we can commit to goes in the contract, in writing, before you build anything on it.
What happens to what you send
Retention, deletion, and whether anything you send is used for training. Ask for all three in writing before the first real request. If the answer we can give is narrower than what you need, you should hear that in the first meeting rather than the fourth.
Capacity that stays warm
A model that is loading can decline a request until it is ready. Startup time depends on the model and serving configuration. On shared capacity that is an occasional refusal. Keeping a model warm for you is a real cost and a real commitment, and it is the thing dedicated capacity actually buys.
Volume, and how it is paid
Committed volume, agreed ahead and invoiced in dinars from a company registered here. No card that works abroad, no foreign exchange file, no annual card ceiling to plan around. For most Tunisian buyers that is the difference between a purchase order and a problem.
Support that is a person
One address, answered by the people who trained the models. What response time we can commit to belongs in the contract rather than on this page, because a number here that nobody has agreed to is worth nothing to you.
Voices you own
Dido clones a voice from one short recording, with consent recorded against it as a condition of the voice existing at all. A brand voice for an IVR or an app is one of the things worth talking about early, because the terms around a cloned voice matter more than the price of one.
What is true today
The parts a sales page would leave out.
Requests made with API keys use only your API credits at the published rates. Subscription allowances do not cover API calls. New accounts receive 10.000 TND in one-time platform starter credit. API credits are purchased separately, starting at 15 TND. Current API rates ↗
API and enterprise pricing
Public API usage follows the current pricing page. Dedicated deployments, support and availability commitments are scoped and quoted separately.
Quality benchmarks
Verified quality benchmarks for the current models are not published yet. We can discuss evaluation on your own text and recordings.
We are small
That is why you get the people who built the models rather than an account manager, and it is also why an availability commitment is something we agree with you rather than something we publish. Both halves of that are worth knowing before you start.
You can test it before you talk to anyone
The models are live. The playground runs them with no key, and a key is one sign-in away. Bring your own audio and your own messages, and judge it on those rather than on this page.
Straight answers
What procurement will ask.
Can it run inside our own infrastructure?
It is one of the shapes we will discuss, and for some buyers it is the only permissible one. It is not something you can switch on today, and anyone who tells you otherwise about a 27B model is selling you a plan rather than a deployment.
Where is our data held?
That is a contract term and we will put it in writing. Ask for it before you send anything real, not after the pilot.
Is anything we send used for training?
Ask, and get the answer in the contract. We are not going to write a promise on a marketing page that your legal team would rightly ignore.
Do you have an SLA?
Not a published one. Availability is agreed per contract, because what we can commit to depends on whether the capacity is shared or yours.
Can you invoice us in dinars?
Yes. That is the ordinary way, not an accommodation.
Why not use a general model that also speaks Arabic?
Try both on your own data before deciding. That is what the playground is for, and it is a better argument than anything on this page. What you will be comparing is a model trained on Derja against one trained on Modern Standard Arabic and asked to cope.