An orchestra of AI models, one conductor.
We're not tied to a single LLM. For every request, ModelMatch™ automatically picks the most efficient model based on complexity, criticality, language, cost — always respecting the privacy tier you've chosen.
Patent PendingOne model per node, not per agent
An agent is a graph of nodes. ModelMatch™ assigns the model to each node, at run time.
Per-node selection
Every node declares its task type: retrieval, analysis, computation, composition, verification. The choice happens there, on every run — not at session start.
Privacy as an upstream filter
The configured tier is an eligibility threshold: only models deliverable at that tier or a stricter one get in. Applied before quality, speed and cost.
Class by criticality
Among eligible models, the class follows error criticality and input ambiguity. Not the perceived importance of the step.
Runtime switch and fallback
Different steps, different makers. Model unavailable or rate-limited → another eligible model, same perimeter, process not interrupted.
Which models we actually use
For every privacy tier, the models deliverable at that tier, sorted by power, speed, cost or ModelMatch eval.
Ranking data from Artificial Analysis
ModelMatch™ Privacy Tiers
| Tier | Name | Description | LLM processing | Included from |
|---|---|---|---|---|
| 0 | Standard SaaS | All LLM providers. Automatic optimization for performance and cost. | Variable, possibly outside the EU (with SCC and no-training clauses) | All plans |
| 1 | EU Resident | Only providers with EU-based servers (Google Vertex AI EU region, Azure OpenAI North Europe, AWS Bedrock EU inference profiles). | Guaranteed in the EU | From Grow |
| 2 | Private Cloud | EU providers on cloud dedicated exclusively to the Customer. Maximum data segregation. | EU, dedicated infrastructure | From Scale |
| 3 | Local LLMcoming soon | Inference on the Customer's own hardware: dedicated machine or VPS running the BlueSky inference engine. | On the Customer's infrastructure | Coming soon |
"EU hosting" refers to the platform core (application, database, storage, backups), always within the European Union. Prompt processing by LLM models follows the privacy level included in your plan.
Sovereign Cloud (AWS European Sovereign Cloud)
Infrastructure in Germany, operated by independent European legal entities with EU-only personnel. Designed for public sector and customers with data sovereignty or CLOUD Act requirements.
Open-weight models on dedicated GPUs
Open-weight models running on GPUs dedicated to the Customer: no model vendor in the data path. The recommended setup for the strictest requirements; cost and sizing tailored per project.
Before a sensitive action, we pause. Always.
When an agent is about to take an action that needs attention, it really stops and waits for your OK — even for days, even after a system restart, without losing the context of where it stopped. It's not a mock-up: it's a technical suspend-and-resume mechanism built into the platform's own architecture.