Content and the work required
Query type and sensitive data can inform which destinations are permitted and which model capabilities the work needs.
How Aventail™ works.
Run supported models on your own hardware. Route each request to the right approved destination. Aventail™ brings inference and orchestration into one system, from local devices and peer machines to servers and frontier APIs.
Aventail™ is an AI orchestration platform combining model deployment, an LLM gateway, a built-in inference engine and policy-based routing.
Workspace or your application sends a request to Aventail™ Link on this machine. Inside the runtime, content, model fit and performance, and device state are evaluated locally. Aventail™ Link applies your rules, then runs or routes the request to one eligible destination: this machine, another approved reachable Aventail™ Link peer, an Aventail™ Link-equipped server or private-cloud or provider machine, or an approved frontier API. Aventail™ Control is outside this request flow and distributes models, configuration and policy through a separate management connection. Frontier APIs execute at their providers.
The request is evaluated where it starts.
Secure on-device content-aware routing happens inside Aventail™ Link. The runtime evaluates query content and policy locally; Aventail™ does not need to receive the query for that analysis. Approved remote destinations still receive requests routed to them.
Model fit and performance
Supported model capabilities, performance and requirements inform suitability. Integration with the inference engine brings model execution conditions into routing decisions.
Device capacity and conditions
The runtime can consider device capacity, performance and health alongside model requirements. Availability, connectivity and resources still determine whether a machine is eligible.
Your organisation’s constraints
Geography, privacy, access and other requirements restrict eligible destinations before preferences such as latency, cost or energy use select among them. No eligible destination means the request cannot be routed under those rules.
Run here. Or use another approved machine.
Aventail™ Link can execute a supported model locally or route the request to a peer: another approved, reachable machine running Aventail™ Link. Model suitability, device conditions and your policy determine which machines can take the work.
Destinations can include your devices, on-premises servers, private-cloud machines and provider compute running Aventail™ Link. Approved frontier APIs run at their providers through your own accounts. Each request executes at one selected destination.
A local model can run offline when it is installed, running and permitted by policy. Registration, downloads, hosted management and routing to remote machines or APIs require connectivity.
Aventail™ Control distributes models, configuration and policies, and manages devices outside the inference request path. Hosted Aventail™ Control receives management, device and routing events. Prompt and response bodies do not pass through hosted Aventail™ Control. Local execution alone does not determine application storage, analytics, plugins or reporting behaviour; review those settings for your deployment.
Supported inference backends
Aventail™ Link officially supports four inference backends. GPU backends run models on a supported graphics processor.
| Backend | Where the model runs |
|---|---|
| Apple Silicon (Metal) | The GPU in Apple Silicon Macs, through Metal. |
| NVIDIA (CUDA) | NVIDIA GPUs, through CUDA. |
| AMD (ROCm) | AMD GPUs, through ROCm. |
| CPU-only | The CPU backend runs supported models on the main processor without requiring a GPU. |
Check the package for your operating system, the model’s format and size, and the memory and storage available on each machine separately.
An LLM gateway and inference engine, together.
Aventail™ Link includes an LLM gateway to approved external AI APIs and machines running Aventail™ Link, alongside an inference engine that runs supported models. It evaluates content, model fit, device conditions and policy locally. Aventail™ Control manages the fleet outside the inference request path.
| Job | Gateway or router | Aventail™ approach |
|---|---|---|
| Connect model endpoints | Unifies access or routes to hosted and self-hosted endpoints. | Routes to approved APIs and Aventail™ Link-equipped machines. |
| Run models | A gateway-only setup relies on a serving layer behind the endpoints. | Aventail™ Link deploys and executes supported models on eligible machines. |
| Choose a destination | Routing capabilities and decision signals vary by offering. | Evaluates query content, model and device conditions, and organisation policy locally. |
| Manage the execution fleet | Deployment and management depend on the serving stack. | Aventail™ Control manages machines, models and configuration outside the inference path. |
A gateway may be enough if your models are already served and you only need endpoint access and routing.
Aventail™ fits when model execution and routing across your own and approved compute should work together.
Start with your workflow.
Review the models, data boundaries and infrastructure you want to use.