Technologies: Which Model Do We Use When?
Huaris AI is vendor-independent. We evaluate model families against the use case, data sensitivity and budget, and base the decision on tests run with your own examples. This page explains when we lead with which model family and what our selection criteria are.
Last updated: September 2026
Model families
The table is a general guide. Version numbers are deliberately left out because models change quickly. Which model works better for a specific task can only be known through an evaluation on your own data.
| Model family | Known for | Typical use | Deployment |
|---|---|---|---|
| Anthropic Claude | Long context, analytical reasoning, coding | Complex document analysis, software development support | Provider API and enterprise plans |
| Google Gemini | Large context window, Google Workspace integration | Organisations on Workspace, multimodal content | Provider API and enterprise plans |
| OpenAI GPT | Rich API and assistant ecosystem | General-purpose assistants, broad integration needs | Provider API and enterprise plans |
| Meta Llama / Mistral | Open-weight models, full control | Cases where data must not leave the organisation | On-premise, private cloud or provider APIs |
Logos and product names belong to their respective owners. Huaris AI does not claim to be an official partner of these companies.
Our selection criteria
When choosing a model we look at these topics together:
- Data sensitivity and regulation: where data is processed, the provider's data use terms, KVKK/GDPR requirements
- Task quality: results on an evaluation set built from your real examples, including performance in Turkish
- Cost: running cost at your usage volume and the assumptions behind the estimate
- Latency and volume: expected response time and number of concurrent users
- Integration: fit with existing tools (for example Google Workspace, the Microsoft ecosystem)
- Vendor dependence: how easy it is to switch to another model when needed
- Operating capacity: whether your organisation can run the model itself
What do we evaluate first, and when?
The following are starting points, not final recommendations. In every scenario we test the candidates on your own data.
| Scenario | Approach we evaluate first | Why? |
|---|---|---|
| Analysis of long contracts and reports | Families known for long context and analytical reasoning | Consistency and capturing detail matter in long texts |
| Team works mostly in Google Workspace | Gemini family | Integration with existing tools eases setup and adoption |
| Need for ready-made assistants and broad third-party integration | GPT family | The API and assistant ecosystem is broad |
| Software development support | Families known for coding ability | Code quality is verified by testing against your own repositories and languages |
| Data must not leave the organisation | On-premise setup with open-weight models | Data stays on infrastructure you control; the quality gap is measured by testing |
Components beyond the model
The quality of an AI solution is not determined by the model alone. Especially in RAG systems, these layers directly affect the outcome:
- Document parsing and chunking
- Embedding models for semantic search
- Search and vector database layer
- The orchestration layer that manages queries and the model
- Evaluation and testing tools
- Logging, monitoring and access control
We choose these components, like the model, to fit your needs, and build them so they can be replaced.
Why we do not tie ourselves to one model
Models and prices change quickly. The model that fits best today may not tomorrow. That is why we build the architecture so that swapping the model does not break the working system, and we keep the evaluation set with you. For details of the method, see the Model Selection and Vendor-Independent Advisory service page, and for general questions the FAQ page.
Let's work together
Let us listen to your processes and goals, and evaluate together where AI can add value.
Send an email