Catalogue
Models catalogue
What we can serve you: what is already loaded, what we start on request, and the country where each model runs.
Compare
The chat models we serve now, with the values we serve: context in tokens, prompt and answer together; the longest answer is max_tokens up to 32,768 on every model. Choose what you need:
All countries 🇨🇭 Switzerland Calls tools Reads images Reasons
| Model | Runs in | Context | Tools | Images | Reasons |
|---|---|---|---|---|---|
gemma-4-26b |
🇨🇭 | 256k | Yes | Yes | No |
ministral-3-14b |
🇨🇭 | 256k | Yes | Yes | No |
mistral-medium-3.5-128b |
🇨🇭 | 256k | Yes | Yes | No |
qwen2.5:1.5b |
🇨🇭 | 32k | — | No | No |
«—» means not tried yet. Every value comes from our engines or from a request through the API, not from the publisher's card.
Ready to use
Already loaded: the first request answers without a start-up wait. There are 6. Each one lists the countries where it runs. A request never leaves the zones allowed for your key (groups of machines, each in one country), and so never leaves their countries.
gemma-4-26b
It reads images as well as text: useful for photos of documents, scans and screenshots. Good at returning structured data from free text, such as extracting fields from a contract or an invoice.
Precision: FP8, 8 bits (Red Hat) · Everything about it
ministral-3-14b
Il modello europeo di Mistral AI, dicembre 2025: legge le immagini, usa gli strumenti e scrive bene in italiano, francese e tedesco. Denso da 14 miliardi di parametri, leggero e veloce.
Precision: As released by Mistral AI · Everything about it
mistral-medium-3.5-128b
Il modello grande di Mistral AI (aprile 2026), denso da 128 miliardi di parametri: legge immagini, usa gli strumenti, ragiona quando glielo si chiede. Il più forte di Mistral a pesi aperti, anche sul codice.
Precision: NVFP4, 4 bits, FP8 in the first and last layers (NVIDIA) · Everything about it
qwen2.5:1.5b
A light, very fast model, for simple and repetitive tasks where response time matters more than nuance.
Precision: 4 bits (Ollama) · Everything about it
bge-m3
Turns a text into a numerical representation that lets you search by meaning instead of exact words. It is the engine of the search in your documents. It works on texts in different languages, even mixed.
bge-reranker-v2-m3
A second pass over the results of a search: it rereads the candidates together with the question and reorders them by actual relevance. It brings the right answer to the top, not just the one that looks most like the words searched for.
If you need a model that is not listed
Ask us: we can add any open-weight model that fits the capacity we have.
What we do not do is resell the closed models of OpenAI or Anthropic: we could not run them ourselves, so we could not make any of the commitments we make on everything else.
Write to support@siati.ai.
Audio
Transcription with whisper-1 and whisper-1-hd; speech with the voices
paola and riccardo in Italian, amy and ryan in
English, thorsten in German, siwis in French. OpenAI voice names are
accepted as synonyms. Details in Audio API.