Which AI models can a Hermes Haven agent use?
Hermes Haven adds a managed routing layer around the model configuration used by Hermes Agent. Model availability is a platform catalogue concern, not a property of the open-source runtime, and it can change as providers and plans change. [A-D6][S-PROXY]
The stable audited statement is narrow: Qwen 3.6 35B is included in every current plan. The audit verified this against both plan configuration and proxy routing. [C13]
What happens when paid model credit runs out?
The proxy implements fallback to Qwen 3.6 when paid credit is exhausted. This is a routing behaviour, not a promise that every request will succeed: availability still depends on the involved services and the request itself. [C16]
| Question | Supported answer | Evidence |
|---|---|---|
| Is there an included model? | Yes. Qwen 3.6 35B is included in every current plan. | [C13] |
| Is fallback implemented? | Yes. The proxy routes to Qwen 3.6 when paid credit is exhausted. | [C16] |
| Is the whole catalogue permanent? | No evidence supports a permanent catalogue. Revalidate it before publication. | Not demonstrated |
| Are all model requests processed in the EU? | No. Premium and customer-configured providers may process requests outside the EU. | [C03] |
See current commercial limits on Pricing.
Where is inference data processed?
The answer depends on the selected route. The public copy must not say that no data leaves the EU: the audited configuration can route premium requests to services such as Ollama Cloud and Gemini, and customer-configured providers may also process data outside the EU. The respective provider's terms and location apply to those requests. [C03]
Persistent agent data and inference payloads are different data flows. A storage statement about the agent workspace must never be used to imply that an external model provider processes prompts in the same location. [C03][C25]
For the broader control picture, read Security.
Does Hermes Haven operate any model services itself?
The audited configuration includes Qwen served on operator-owned GPU hardware. That supports “operator-owned hardware”, but not “European GPU cluster” or a specific data-centre location. [C04]
Speech-to-text also runs on operator-owned GPU hardware without a per-minute external API charge; see Integrations. [C23]
Can a user connect a different model provider?
The audited product copy and proxy configuration support premium routes and user-configured providers. Because the exact provider and model catalogue is time-sensitive, the publishable page should populate a live or versioned catalogue rather than hard-code an exhaustive list here. [C03]
Before sending sensitive prompts to any route, verify:
- the provider handling that model;
- the provider's processing region and retention policy;
- the plan entitlement or credit required;
- the fallback behaviour that applies.
What must be reviewed before this page is published?
- Replace any dynamic catalogue placeholder with a dated export from the current catalogue.
- Do not claim universal EU-only processing. [C03]
- Do not describe operator-owned hardware as a European data-centre cluster. [C04]
- Keep the Qwen inclusion and fallback statements synchronized with plan and proxy configuration. [C13][C16]
Learn how the operational layer fits around models on Platform.