Legal & Security
Transparency Center
Which models power Simit GPT, how data is used, and what our limitations are.
Which models are used?
| Component | Model | Status |
|---|---|---|
| Simit Chat (general) | Ready-made API models (third-party infrastructure) | Live |
| DeepThink mode | The reasoning mode of ready-made API models | Live |
| Simit-TR | Local Turkish model fine-tuned from an open-weight base model | Planned |
| Türkiye Knowledge Base | Retrieval-augmented generation (RAG) on PostgreSQL + pgvector | Planned |
Our "homegrown" goal is a journey: Simit's interface, Model Router, system prompt and all infrastructure are built by us. At the language model layer we use ready-made API models today while developing our own model, Simit-TR. Details: Simit-TR, Hybrid Model Routing
How is Simit set up to behave?
Simit's system prompt is based on these principles:
- Answer in the user's language, with natural and correct Turkish first
- Keep Türkiye's context in mind when relevant
- Never pretend to know what it doesn't; flag information that may be outdated
- Refer legal, medical and financial questions to experts and official sources
Data
- Your messages are processed to generate answers; message content is not written to server logs.
- Guest chats stay in your browser; members' chats are stored in their account.
- Conversations are not used for training without your permission.
- Details: Privacy Policy
Known limitations
- The model's knowledge freezes at a date; Simit doesn't search the web yet.
- Answers can occasionally be wrong or made up ("hallucinations").
- Earlier details can be missed in long conversations.
- Simit is not a person and has no feelings or personal experiences.
Feedback: info@cmtedu.com