Provider setup

Set up China-hosted models

DeepSeek / Qwen / Moonshot — setup mirrors OpenAI.

Overview

Major Chinese model vendors all publish OpenAI-compatible endpoints — adding one in Kition is a custom provider with wire responses, just swap baseURL + key. Sections below list official baseURLs and the common model ids per vendor.

Network notes: hitting OpenAI / Anthropic from mainland China often times out; direct endpoints from Chinese vendors are low-latency without a proxy — noticeably better for long-running agent jobs.

DeepSeek

baseURL: https://api.deepseek.com (no /v1 suffix)
Auth: Authorization: Bearer sk-...
Models: deepseek-chat (V3 general), deepseek-reasoner (R1 reasoning)
Cheap, 64k context; reasoner returns a reasoning_content field that Kition renders automatically

Qwen / DashScope

Alibaba Cloud DashScope publishes both a native protocol and an OpenAI-compatible mode. Kition uses the compatible mode; note the baseURL ends in /compatible-mode/v1.

{
  "name": "qwen-dashscope",
  "baseURL": "https://dashscope.aliyuncs.com/compatible-mode/v1",
  "apiKey": "sk-...",
  "wire": "responses",
  "defaultModel": "qwen3-max"
}

Moonshot / Kimi

baseURL: https://api.moonshot.cn/v1
Models: kimi-k2-thinking (reasoning), kimi-latest (default), moonshot-v1-128k (long context)
Long-context is its strength — great for dumping a whole vault in for research

Zhipu / GLM

baseURL: https://open.bigmodel.cn/api/paas/v4
Models: glm-4.6, glm-4.6-flash
Key format is <id>.<secret> — paste the whole string into apiKey

Doubao / Volcengine

baseURL: https://ark.cn-beijing.volces.com/api/v3
The model field takes an endpoint id (e.g. ep-2025...), not the model name — created in the Volcengine console
Auth: Authorization: Bearer <ark-key>

Baichuan & Yi

Baichuan: https://api.baichuan-ai.com/v1, model Baichuan4-Turbo
Yi (01.AI): https://api.lingyiwanwu.com/v1, model yi-lightning / yi-large
Both use standard Bearer auth

Common errors

401 — wrong key; Chinese vendors often use prefixes (sk-, xx.yy) — do not miss characters
404 model_not_found — vendor renamed the model (e.g. qwen-max is now qwen3-max)
429 — per-key QPS is low by default; raise it in the vendor console
baseURL path wrong — DeepSeek omits /v1, DashScope uses /compatible-mode/v1; check carefully
Content moderation block — Chinese vendors filter prompts; a 400 means rephrase or switch model

← Back to help center More in Provider setup →

Custom OpenAI-compatible provider

Any endpoint speaking the Responses shape — self-hosted Ollama / vLLM / company gateway.