Documentation
Choose between China-developed and global models in the NexoRouter documentation.
Choose between China-developed and global models
NexoRouter exposes China-developed and global model families through the same OpenAI-compatible interface. The right choice depends on the workload, not the country label.
Use the current catalog and a controlled test set to compare quality, cost, latency, and availability before moving production traffic.
Start with verified data
For every candidate:
- Confirm it appears in Models or
GET /v1/models. - Copy the exact NexoRouter model ID.
- Open the model profile and review its billing unit, published rates, capabilities, and official vendor source.
- Confirm the API key scope allows the model.
- Run the same small test set and inspect Usage Logs.
Do not rely on an old spreadsheet, screenshot, social post, or vendor alias. Model IDs, features, availability, and rates can change.
Build a useful shortlist
| Workload | What to compare |
|---|---|
| Customer support in Spanish | Instruction following, Latin American Spanish, latency, output cost |
| Extraction and classification | Structured consistency, short-output cost, error rate |
| Coding and agents | Tool-call behavior, context needs, retry rate, completion quality |
| Long documents | Context limit, input price, timeout behavior |
| Image, video, or audio | Endpoint, input format, billing unit, asynchronous behavior |
China-developed families such as DeepSeek, Qwen, Kimi, and GLM can add useful price and capability options. They are not automatically cheaper, safer, or better for every workload. Compare the published NexoRouter rates and the result of your own test set.
Query the current catalog
curl https://api.nexorouter.com/v1/models \
-H "Authorization: Bearer $NEXOROUTER_API_KEY"
Keep the selected ID in configuration:
NEXOROUTER_MODEL=exact-model-id-from-the-catalog
Model IDs are case-sensitive.
Run a repeatable evaluation
Create 20 to 50 representative prompts. Include normal requests, ambiguous requests, long input, and expected failures.
For each candidate, keep these values constant:
- System instruction.
- User prompt.
temperature.max_tokens.- Timeout and retry policy.
Record:
- Task success.
- Response quality.
- Input and output tokens.
- Cost shown in Usage Logs.
- Latency and status.
- Request ID for failed calls.
Avoid declaring a winner from one prompt.
Promote safely
- Validate the integration with a low-cost model.
- Test the shortlist against the same prompts.
- Create a dedicated production key with a budget and model scope.
- Send a small percentage of traffic first.
- Monitor Usage Logs and error rate.
- Keep a reviewed fallback model only when your application can tolerate different output behavior.