use-case
Model originators, hosted clouds, aggregators, and local runtimes
Direct answerA provider guide answers who develops the model; a channel answers where a request is routed. The same model through another host can change authentication, protocol, region, version, and data boundary.
Updated · Reviewed
Beginner: who built it versus where it is called
An originator defines weights, capability, and lifecycle. A hosted cloud offers third-party models in its regions and identity system. An aggregator exposes several suppliers through a common protocol. A workflow platform exposes an application or knowledge base. A local runtime loads models on owned hardware. An account bridge converts another identity system into access. All may appear in channel configuration, but they are not all model vendors.
Six integration shapes
| Type | Examples | Verify |
|---|---|---|
| Originator | OpenAI, Anthropic, Google, DeepSeek | Official ID, native protocol, retirement |
| Hosted cloud | AWS Bedrock, Azure AI Foundry, Vertex AI | Region, deployment name, platform version, origin |
| Aggregator | OpenRouter, SiliconFlow, Replicate | Effective supplier, routing, protocol subset |
| Workflow | Dify, Coze, FastGPT | App ID, conversation state, knowledge semantics |
| Local runtime | Ollama, Xinference | Artifact, license, hardware, network exposure |
| Account bridge | OAuth or account-based access | Identity, scopes, session, service terms |
Minimal verification
Query /v1/models with the site token, then inspect the marketplace for exact ID, enabled endpoint, and price. A platform name inside an ID does not establish model authorship, and a platform catalog does not guarantee every region, account, or endpoint.
A channel switch is not a base-URL edit
Retest authentication, model or deployment name, fields, tool schema, stream events, errors, usage, media delivery, and limits. A host may pin an older snapshot, an aggregator may fail over among suppliers, and a local runtime changes with quantization, templates, and hardware.
Data and compliance boundary
Retain the channel and originator for every request. Determine which entities process data, in which region, for how long, and for what abuse monitoring. Automatic failover can change that boundary. A zero-retention requirement cannot rely on a model name or a platform-wide slogan.
Expert routing governance
Give every channel a capability allowlist, health score, cost, latency, data class, and failover policy. Do not fail over every error: validation, permission, balance, and policy failures rarely improve on another supplier. After failover, keep Request-ID, effective supplier, model snapshot, and settlement data for reproduction and accountability.
Use cases
- Choose channels while preserving model attribution
- Govern multi-supplier routing and data boundaries
API protocols
/v1/models
FAQ
Are OpenRouter, AWS, or Azure model originators?
They primarily aggregate or host third-party models. Capability belongs to the originator; the platform guide describes hosting and protocol boundaries.
Should the same model be retested after changing channel?
Yes. Snapshot, region, parameter mapping, stream events, limits, data policy, and price can differ.
Why are there more channel choices than provider guides?
Channels also include aggregators, workflows, local runtimes, compatibility proxies, and account bridges that do not own a model family.
Related guides
Official sources
- Amazon Bedrock Supported Models Official
- Azure AI Foundry Models Official
- OpenAI API Reference Official
兔子API