Connect AWS and Chinese model families

AWS onboarding selects Bedrock models already registered in the Ani Router catalog. Choose the same region where your account has model access; the guide offers common US, Europe, Singapore, and Tokyo regions.

Onboard and verify

1. 1Select AWS in Admin → Provider accounts, then choose the region and authentication mode.
2. 2Enter the required identity fields and select only models this account can invoke.
3. 3Save to create the account and selected model bindings.
4. 4Run a connection test for every binding in Model availability.
5. 5Review pricing and team access before enabling production routing.

AWS Bedrock model access

Model Studio uses an OpenAI-compatible regional endpoint. Ani Router gives each built-in model a stable alicloud/<upstream-model-id> name, while the account binding supplies the endpoint and encrypted API key.

What you need

• A unique account name and the Model Studio region where model access is enabled.
• A Model Studio API key for that account and product entry.
• A Workspace ID for Singapore, Beijing, Hong Kong, Tokyo, or Frankfurt.
• At least one built-in model that is actually enabled in your Model Studio console.

Catalog and production checks

• Built-in families include Qwen, DeepSeek, Kimi, GLM, and selected MiniMax models.
• The catalog is a versioned snapshot, not live discovery. Deselect anything your current console has not enabled.
• New unpriced models stay inactive. Enter reviewed pricing, enable the model, test the binding, and then add team routing.
• Existing model pricing and activation are preserved when onboarding is repeated.

Create a Model Studio API key · Model Studio model list

Use TokenHub for new integrations. Ani Router stores built-in models under stable tencent/<upstream-model-id> names and derives the default endpoint from the selected region.

Onboard and verify

1. 1Select Tencent Cloud in Provider accounts and choose Guangzhou or Singapore.
2. 2Enter an API key issued for the endpoint you are using; keep the TokenHub default for new accounts.
3. 3Select only the Hunyuan, DeepSeek, GLM, Kimi, MiniMax, or Qwen models enabled in your console.
4. 4Save, review every zero-price warning, and keep unreviewed models inactive.
5. 5Test each binding before configuring team routes and allowlists.

TokenHub models and protocols · TokenHub migration guide

InternLM, Yi, Baichuan, Xiaomi MiMo, and any family missing from the Alibaba or Tencent built-in catalog must not be assumed to be hosted by those clouds. Deploy an open-weight model yourself—or use a vendor that gives you a dedicated OpenAI-compatible endpoint—then register it as a manual upstream.

Manual onboarding path

1. 1Deploy the model with vLLM, SGLang, Ollama, or another OpenAI-compatible server and verify its exact upstream model ID.
2. 2In Admin → Models, create an OPENAI_COMPATIBLE model with a unique display name and exact upstream model ID. Leave model-level credentials empty when the account pool will own them.
3. 3In Admin → Provider accounts, create an OTHER account with the HTTPS base endpoint and provider API key.
4. 4In Model availability, bind the account to the model and set an upstream model override when it differs from the display name.
5. 5Use the default Chat Completions format first. A different API format needs reviewed model metadata and a successful compatibility test.
6. 6Enter reviewed prices, test non-streaming and streaming requests, configure team access, and only then enable production traffic.

vLLM OpenAI-compatible server

Production launch checklist

A saved provider account is only the first step. Complete all checks below before allowing team traffic.

• The provider, region, endpoint, and credential belong to the same cloud product entry.
• Every selected model is enabled for the account in the current provider console.
• Input and output prices are reviewed; no zero placeholder is treated as free usage.
• The model and account-model binding are active and pass a connection test.
• Streaming and non-streaming calls work with the exact namespaced or custom model ID.
• Team routes, allowlists, budgets, and rate limits are configured.
• Provider credentials are stored only in Admin or deployment secrets and never committed to Git.