Persian large language models on Iranian infrastructure
Enterprise AI in Iran faces two obstacles: models that do not understand Persian well, and services that take data abroad or need a VPN. Miras solves both with its own infrastructure.
Three models, one gateway, inside Iran
Persian large language model
Open-weight language models hosted on Miras infrastructure and tuned on Iranian industry and construction data: enterprise assistant, technical knowledge search, report summarisation, letters and minutes in Persian.
Vision-language model
A model that sees an image and explains it in Persian: site photos, handwritten invoices, drawings and industrial equipment become text and structured data.
Persian speech-to-text
Miras Persian speech recognition for site voice reports, meetings and call centres; familiar with industry vocabulary and colloquial speech.
A dedicated GPU server in Tehran
Miras has invested in its own compute: a dedicated server with an NVIDIA A100 GPU in Tehran and a single model gateway for all group products. Models are fine-tuned, tested and served in Tehran.
- Data stays in Iran. By default no data is sent to a foreign cloud service; a foreign fallback model is used only with the customer’s written consent.
- No VPN. Every product works on Iranian mobile data.
- Open models, your ownership. We work on open-weight models; the model tuned for your organisation can be delivered to your own server.
- Scalable. From one server inside your organisation to Miras infrastructure for several systems; each system has its own key and usage cap.
What you receive
Miras assistants already work inside the Miras PM portal and the group’s products; the same infrastructure is available to your organisation.
We use our models first ourselves
Local model first
Every request goes to Miras’s own models first; a paid foreign service is a fallback only, and only with the customer’s written consent.
Correct Persian
Persian digits, the Persian calendar, the Vazirmatn typeface; model output is reviewed by a native Persian speaker.
Text is the record, voice is an input
Speech becomes editable text and nothing is sent without the user’s confirmation.
Controlled cost
Each system has its own key and usage cap; no paid model runs in an automatic loop.
Frequently asked questions
Which Persian language model is best for an organisation?
The one tuned on the organisation’s own data and run on infrastructure that keeps the data in Iran. Miras hosts strong open-weight models on its GPU server in Tehran and fine-tunes them for each organisation with its own documents and terminology.
What does Iranian AI infrastructure mean for us?
A compute server inside the country, a deliverable model and no dependency on a foreign cloud. Miras has this infrastructure and offers it as a service or as a deployment on your own server.
Can we have the model on our own server?
Yes. The tuned model, gateway and documentation are deployed on your server, and the Miras team maintains and updates it.
Does your Persian speech-to-text handle accents and industrial vocabulary?
The Miras speech model is built for site and industry voice reports; in the pilot its accuracy is measured on your own team’s voices and tuned where needed.
Bring intelligence to your plant or project
We start with a data audit: where your data is, where cost leaks, and which model pays back first. Tehran · Isfahan · Noshahr · Dubai.