兔子API
provider
Baidu ERNIE
Direct answerDirect Qianfan access uses https://qianfan.baidubce.com/v2 and a Bearer bce-v3/... key. Preset models use the documented model parameter; custom deployments use a service API name, while this gateway uses its own token.
Updated · Reviewed
Qianfan versus ERNIE
The Qianfan catalog includes Baidu ERNIE and partner models, so not every listing belongs to ERNIE. Verify author, model parameter, and capability matrix. This gateway page declares only /v1/chat/completions; native Qianfan embedding, image, and other service APIs are not gateway routes by implication.
Direct access versus this gateway
Qianfan's V2 OpenAI-compatible base is https://qianfan.baidubce.com/v2, with Chat at https://qianfan.baidubce.com/v2/chat/completions, authenticated by a Bearer bce-v3/... API key. This gateway uses its own base URL, Bearer token, and enabled model name. Do not reuse stale V1 tutorials or send a Baidu key to this site.
Preset models versus custom services
A preset model uses the model parameter published in the model list, such as ernie-3.5-8k; a trained deployment or custom service uses its service API name. Console display names are for humans. Through this gateway, a channel may map an enabled gateway name to either value, so clients use only the public gateway name.
appid, streaming, and tools
An optional appid header supports usage and billing attribution, and an API key can be scoped to all resources or one appid. Text generation supports streaming and tools; model-generated function arguments still require application validation and execution. Response headers expose request and input/output token limits and remaining capacity for proactive throttling.
Errors, limits, and cost safety
A 429 can separately indicate RPM, aggregate TPM, input/output TPM, application QPS, or cluster limits, so combine the business code with rate-limit headers. Coding Plan has a dedicated key and must not be mixed with pay-as-you-go. Cap output, concurrency, and budget, alert on usage, retain request IDs, and use the current Qianfan pages for price and model status.
Reviewed provider catalog
Latest models by capability
These are provider-catalog models, not a promise of availability on this site. Confirm callable IDs, endpoints, and pricing in the model marketplace.
Text, reasoning, and code
-
ernie-5.1StableBaidu ERNIE's current flagship text model for complex reasoning, multimodal understanding, and agents.
Image generation and editing
-
musesteamer-air-imageStableBaidu's current MuseSteamer Air model for image generation and editing.
Video generation and editing
-
musesteamer-air-i2vStableBaidu's current MuseSteamer Air image-to-video model.
-
musesteamer-2.1-turbo-i2vStableThe high-quality image-to-video variant in the MuseSteamer 2.1 family.
Speech and audio
-
audio-realtime-nearPreviewBaidu's Pro near-field end-to-end speech-language model for realtime text and audio output, currently in public beta.
-
audio-realtime-farPreviewBaidu's Pro far-field end-to-end speech-language model for distant capture, currently in public beta.
Embeddings, retrieval, and reranking
-
embedding-v1StableBaidu's text embedding model for semantic search, similarity, and RAG.
Use cases
- Chinese understanding and generation
- Reasoning and multimodality
- Enterprise and industry workloads
API protocols
/v1/chat/completions
FAQ
What base URL and API key does Qianfan V2 use?
For a direct OpenAI SDK, use https://qianfan.baidubce.com/v2; Chat is /v2/chat/completions, authenticated with a Bearer bce-v3/... key. Gateway calls instead use this site's base URL and token, never the Baidu key.
Should model contain an ERNIE display name or a custom service name?
A preset model uses the model parameter published in the Qianfan model list, such as the official ernie-3.5-8k example. A trained deployment or custom service uses its service API name. A console display title replaces neither.
Is appid required, and what does it do?
appid is not the primary authentication key; it is an optional header for application-level usage and billing attribution. API keys can also cover all resources or a specific appid, so production workloads should separate appids and keys for accountable billing.
How do I tell whether a Qianfan 429 is RPM or TPM?
Read the provider business code and response headers that report request, input-token, and output-token limits and remaining capacity. A 429 can also come from application QPS or cluster limits. Coding Plan uses a dedicated key that cannot be mixed with pay-as-you-go.