Skip to main content

Baidu Qianfan Provider Guide

Baidu Qianfan is a Tier-2 catalog provider: its integration is one JSON file (src/lib/providers/catalog/baidu-qianfan.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs-verified only, not yet live-verified. An unauthenticated GET https://qianfan.baidubce.com/v2/models on 2026-09-29 answered HTTP 403, so the model ids come from Baidu's public docs and the roster has not been checked. No account was created and no API key was used to build this entry. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: baidu-qianfan (aliases qianfan, baidu)
  • Protocol: OpenAI-compatible (/chat/completions). The quickstart says "千帆ModelBuilder提供了与OpenAI兼容的使用方式，用户只需调整api_key、base_url、model等参数，就可以通过OpenAI SDK调用千帆ModelBuilder推理服务。" and the Inference API V2 page says the V2 model-service API is "完全兼容OpenAI标准（包含身份认证、接口协议）"
  • Base URL: https://qianfan.baidubce.com/v2
  • Default model: ernie-5.1
  • Models in catalog: 12, taken from the vendor's model-list page (updated 2026-09-16, retrieved 2026-09-29)
  • Streaming: supported — the structured-outputs page shows an OpenAI SDK example with stream=True that iterates over the chunks
  • Tool calling: model-dependent — the function-calling page (updated 2026-07-09) documents an OpenAI-style tools parameter and lists ernie-5.1 and ernie-5.0 in its 支持模型范围 (supported models) section
  • Tools while streaming: not declared (false) — the function-calling page's request examples send "stream": false
  • Structured output: true on this entry — the structured-outputs page (updated 2026-05-27) documents the response_format field with text, json_object and json_schema, has a json_schema object description containing "只有部分模型支持" (translation: only some models support it), and lists DeepSeek-V4-Pro and GLM-5.1, among others, in its 支持模型范围 (supported models) section
  • Structured output + tools together: not declared (false) — no combined probe was possible without credentials
  • Embeddings: not declared on this catalog entry (embeddings is false). The model-list page has a 文本向量 (text embedding) section; this flag describes the NeuroLink entry, not the vendor's own API surface
  • Thinking: not declared (false) — the deep-thinking page documents thinking and enable_thinking request parameters for the models it lists and a reasoning_content output field; this entry declares no thinking capability
  • Billing: free-tier — see Billing below
  • Key format: none declared (apiKeyFormat is null), so NeuroLink does not validate the key
  • Scheduled retirements: the model-retirement page (updated 2026-09-24, retrieved 2026-09-29) has rows with 模型退役日期 "2026-10-29" for ERNIE-5.0, ERNIE-4.5-Turbo-128K, ERNIE-4.5-Turbo-32K, ERNIE-4.5-Turbo-VL-128K and ERNIE-4.5-Turbo-VL-32K

Quick Start​

1. Get an API key​

  1. Visit: https://console.bce.baidu.com/qianfan/ (an unauthenticated GET on 2026-09-29 answered HTTP 200 with the page title 百度智能云千帆大模型平台; sign-in was not attempted while building this entry)
  2. Create an API key: the API KEY page (https://cloud.baidu.com/doc/qianfan/s/wmh8l6tnf, updated 2026-04-14) says to log in to the 百度千帆 platform, click 系统管理 in the left navigation, open API Key, then click 创建 API Key and enter a name; the quickstart (https://cloud.baidu.com/doc/qianfan/s/rmh4stn9m) links 控制台-安全认证-API Key to https://console.bce.baidu.com/iam/#/iam/apikey/list
  3. Authentication is HTTP Bearer: the quickstart (https://cloud.baidu.com/doc/qianfan/s/rmh4stn9m) says the Authorization request header must carry the API key with Bearer in front of it; NeuroLink does not validate the key format
  4. Billing as the vendor states it: the new-user free-quota page (https://cloud.baidu.com/doc/qianfan/s/Imi2rpirg, updated 2025-11-17) says that after you visit the platform and agree to the user agreement the system activates it automatically and issues a new-user free token quota; its table has 17 rows under 服务名称, 赠送Tokens量 and 有效期, and the page says the quota applies to online inference calls of preset model services, not batch inference. The billing page (https://cloud.baidu.com/doc/qianfan/s/wmh4sv6ya, updated 2026-09-24) lists text-generation prices under "按量后付费" in 元/千tokens, for example "ERNIE 5.1" (版本名称 "ERNIE-5.1") with 在线推理 "0.004" for 子项 输入（输入<=32k） and "0.018" for 子项 输出（输入<=32k）
  5. Set BAIDU_QIANFAN_API_KEY in your .env file

2. Configure​

export BAIDU_QIANFAN_API_KEY=your-api-key
export BAIDU_QIANFAN_MODEL=ernie-5.1   # optional — overrides the default model
export BAIDU_QIANFAN_BASE_URL=https://qianfan.baidubce.com/v2   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "baidu-qianfan",
  model: "ernie-5.1",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider baidu-qianfan

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "baidu-qianfan",
  credentials: { baiduQianfan: { apiKey: process.env.BAIDU_QIANFAN_API_KEY } },
});

Billing​

The new-user free-quota page (updated 2025-11-17) says that from 2025-10-24 00:00, after you visit the platform and agree to the user agreement, the system activates Qianfan automatically and issues a new-user free token quota. Its table has 17 rows, and the 服务名称 column reads: ERNIE-4.5-Turbo-128K, ERNIE-4.5-Turbo-32K, ERNIE-4.5-Turbo-VL, ERNIE-X1-Turbo-32K, DeepSeek-R1, DeepSeek-R1-250528, DeepSeek-V3-250324, DeepSeek-V3.1-250821, DeepSeek-V3.1-Think-250821, Kimi-K2-Instruct, Qwen3-235B-A22B-Instruct-2507, Qwen3-30B-A3B-Instruct-2507, Qwen3-Coder-30B-A3B-Instruct, Qwen3-Coder-480B-A35B-Instruct, bge-large-en, bge-large-zh and qianfan-sug-8k. In rows 1 to 17 the 赠送Tokens量 cell reads "100万" and the 有效期 cell reads "3个月". The page says the quota applies to online inference calls of preset model services, not batch inference. The entry records free-tier from that page.

The model-retirement page (updated 2026-09-24, retrieved 2026-09-29) has rows with 退役模型版本 "DeepSeek-V3.1-250821" (模型退役日期 "2026-06-30"), "Kimi-K2-Instruct" (模型退役日期 "2026-03-26") and "ERNIE-4.5-Turbo-128K" (模型退役日期 "2026-10-29").

The billing page (updated 2026-09-24) lists text-generation prices under "按量后付费" in 元/千tokens (yuan per thousand tokens). For example, its row with 模型名称 "ERNIE 5.1" and 版本名称 "ERNIE-5.1" has 在线推理 "0.004" for 子项 输入（输入<=32k）, "0.018" for 输出（输入<=32k）, "0.006" for 输入（32k<输入<=128k） and "0.022" for 输出（32k<输入<=128k）. The catalog entry carries no pricingPerMTok; the billing page states prices in 元/千tokens.

Models​

ModelContextVisionMax outputNotes
ernie-5.1 ⭐128kno65,536NeuroLink default. Listed under the "ERNIE系列-旗舰模型" heading; the "推荐模型" table cell begins "ERNIE 5.1是文心系列最新模型". Listed on the function-calling page.
ernie-5.0128kyes65,536Listed in the vendor's 视觉理解 (vision understanding) ERNIE table and on the function-calling page; described as "文心5.0是原生全模态大模型". The model-retirement page has a row with 退役模型版本 "ERNIE-5.0" and 模型退役日期 "2026-10-29".
ernie-4.5-turbo-128k128knonot setListed under the "ERNIE系列-旗舰模型" heading and in the "推荐模型" table; the page states max output as "16k" in one table and "[2，12288]" in another, so the entry leaves it unset. The model-retirement page has a row with 退役模型版本 "ERNIE-4.5-Turbo-128K" and 模型退役日期 "2026-10-29".
ernie-4.5-turbo-32k32kno12,288Listed under the "ERNIE系列-旗舰模型" heading. The model-retirement page has a row with 退役模型版本 "ERNIE-4.5-Turbo-32K" and 模型退役日期 "2026-10-29".
ernie-4.5-turbo-vl128kyes16,384Listed in the vendor's 视觉理解 (vision understanding) ERNIE table; models.visionModel. The model-retirement page has a row with 退役模型版本 "ERNIE-4.5-Turbo-VL-128K" and 模型退役日期 "2026-10-29", and a row with 退役模型版本 "ERNIE-4.5-Turbo-VL-32K" and 模型退役日期 "2026-10-29".
deepseek-v4-pro1Mno393,216NeuroLink fallback. The "推荐模型" table cell begins "核心定位：超强通用旗舰大模型，Agent能力大幅提高。"; listed on the structured-outputs page and in the deep-thinking page's thinking list.
deepseek-v4-pro-08131Mno393,216Listed in the DeepSeek tables of the text-generation and deep-thinking sections.
deepseek-v4-flash-07311Mno393,216Listed in the DeepSeek tables of the text-generation and deep-thinking sections.
glm-5.31Mno131,072Listed under "其他" in the 深度思考 (deep thinking) section.
glm-5.21Mno131,072Listed under "其他" in the 深度思考 (deep thinking) section.
glm-5.1198kno131,072NeuroLink fallback. Listed under "其他" in the 深度思考 section, on the structured-outputs page and in the deep-thinking page's thinking list.
qwen3.5-397b-a17b256kno65,536NeuroLink fallback. Listed under "Qwen系列" in the 深度思考 (deep thinking) section.

The figures are read from the model-list page (updated 2026-09-16, retrieved 2026-09-29). The Context column shows the page's strings ("128k", "1M" and similar); the catalog's contextWindow values are NeuroLink's numeric entries for them, converted with k = 1,000 and M = 1,000,000. Max output is the top of the range the page states, for example "[1，65536]". The deep-thinking section's note says "以下最大输出长度为包含思维链输出的总长度" (translation: the max output lengths below are the total including chain-of-thought output), which applies to the glm-* and qwen3.5-* rows. Vision is yes for the models the page lists in its 视觉理解 (vision understanding) section. The vision-understanding page (docs) documents image input as OpenAI-style image_url content parts.

models.defaultContextWindow (32,000) and models.defaultMaxOutputTokens (4,096) are placeholders the vendor does not publish; they apply to a model id that is not in the catalog. Each per-model contextWindow is a placeholder too: NeuroLink's conversion of the page's "128k", "1M" style strings.

Fallback order when the default is unavailable: deepseek-v4-pro → glm-5.1 → qwen3.5-397b-a17b. The runtime fallback model name the loader derives (fallbacks[1]) is glm-5.1.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Baidu Qianfan — docs-verified only, not live-verified:

ProbeResult
Rosterunauthenticated GET https://qianfan.baidubce.com/v2/models, HTTP 403, JSON body with code "AccessDenied" and message "Access denied.", 2026-09-29 — no key used, so the 12 model ids come from the model-list page and the roster has not been checked
Auth-failure shapeunauthenticated GET https://qianfan.baidubce.com/v2/chat/completions, HTTP 401, JSON body with error.code "invalid_iam_token", 2026-09-29. The error-code page (updated 2026-04-22) lists 错误码 "invalid_iam_token" as HTTP 401 with 错误信息 "IAM Certification failed" and rows of 类型 "rate_limit_exceeded" as HTTP 429
Model retirementspublic model-retirement page (updated 2026-09-24), read 2026-09-29 — rows with 模型退役日期 "2026-10-29" for ERNIE-5.0, ERNIE-4.5-Turbo-128K, ERNIE-4.5-Turbo-32K, ERNIE-4.5-Turbo-VL-128K and ERNIE-4.5-Turbo-VL-32K
Billingpublic free-quota page (updated 2025-11-17) states a new-user free token quota issued automatically after agreeing to the user agreement, 2026-09-29
Tools / structured outputdocumented on the function-calling and structured-outputs pages respectively; neither was exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=baidu-qianfan with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

SymptomCauseFix
HTTP 401 with code invalid_iam_tokenThe error-code page says the bearer token is invalid or expired, or the request's Authorization header did not add BearerThe quickstart says to put Bearer in front of the API key in the Authorization header; NeuroLink reads BAIDU_QIANFAN_API_KEY
HTTP 429 with 类型 rate_limit_exceededThe error-code page lists 错误码 rpm_rate_limit_exceeded (错误信息 "Rate limit reached for RPM") and tpm_rate_limit_exceeded ("Rate limit reached for TPM")The same page's 说明 for those rows includes "适当控制请求频率" (translation: control the request frequency appropriately)

See also​