Security and models
AI models
The model is a replaceable execution component — Slimaco matches it to the task and isn't locked to a single provider.
Slimaco is not a competing LLM. The platform uses models as execution services and can choose a model or model profile for the type of task at hand. The LiteLLM layer provides neutral model aliases, and the orchestration logic can route tasks by difficulty, cost, speed and tool-use capability.
Where models are processed
Company knowledge is stored in the EU, and searching it (embeddings) runs on Slimaco's own infrastructure without going to an external provider. Only the question and the knowledge fragments needed for the answer reach the language model, with personal data masked. Some models run in the Microsoft Azure EU data zone with an EU-only processing guarantee, including the default model, GPT-5.4 Mini. The other selectable models process data globally.
| Model | Infrastructure | Processing | EU only |
|---|---|---|---|
| GPT-5.4 Mini (default model) | Microsoft Azure | EU data zone (Poland, Sweden, France) | Yes |
| DeepSeek V4 Flash | Microsoft Azure | EU data zone (Poland, Sweden, France) | Yes |
| Mistral Large 3 | Microsoft Azure | EU data zone (Poland, Sweden, France) | Yes |
| GPT-5.4 Nano | Microsoft Azure | Global | No |
| GPT-6 Astra | Microsoft Azure | Global | No |
| DeepSeek V4 Pro | Microsoft Azure | Global | No |
| Kimi K2.6 | Microsoft Azure | Global | No |
| Claude Sonnet 5 | Azure AI Foundry (Sweden endpoint) | Global | No |
| Claude Opus 5.5 | Azure AI Foundry (Sweden endpoint) | Global | No |
| Claude Haiku 4.5 | Azure AI Foundry (Sweden endpoint) | Global | No |
| multilingual-e5-large (knowledge search) | Slimaco infrastructure | Local, no external provider | Yes |
As of October 4, 2026. The model list changes as the platform evolves.
What depends on the model, and what stays with Slimaco
| Depends mainly on the model | Stays the responsibility of Slimaco and the organization |
|---|---|
| Generation and reasoning quality | The scope of approved knowledge |
| Speed and cost of a single call | Permissions to documents and systems |
| Supported context length and modalities | Versioning, review and content owners |
| Tool-calling capability | The list of tools available for a given task |
| Multilingual quality | Consent gates, audit and human accountability |
| Robustness to complex instructions | The rollout process, adoption and measuring impact |
How to compare LLMs
- Quality on your own Golden Set, not on a generic benchmark.
- Accuracy of answers grounded in company documents.
- Correctness of citations and the ability to decline when evidence is missing.
- Handling of Polish (or the client's language) and domain terminology.
- Tool calling: argument correctness, resilience to errors, and behavior on refusal.
- The cost of the whole task, not just the token price: retrieval, reranking, follow-up steps, corrections and human time.
- Answer latency and stability under load.
- Data-processing terms, region, provider retention and contracts.
- Availability of the model version and a migration strategy once it changes.
BYO AI
Slimaco offers a BYO AI mode: the ability to use the client's own agreed access to models, or a separate billing profile. This is not the same as an on-premise deployment of the whole platform — BYO AI concerns the model layer, and the platform's deployment model is decided separately.
Deployment model
Slimaco AI runs primarily as a SaaS service, but the platform can also be deployed on-premise or in dedicated client infrastructure — both modes are available, and the exact model, processing region and way of integrating with AI models are agreed individually as part of the deployment conversation and contract.