Chinese LLM API setup: DeepSeek, Qwen, Kimi, Doubao and Zhipu

Bottom line: When accessing domestic LLM, do not guess the model ID based on the manufacturer name. The correct process is to first query the real-time model list, then use the real model ID to complete the minimum request, and finally verify business capabilities such as streaming output, tool invocation, image or retrieval item by item.

If your question is how to choose or unify access to Chinese LLM APIs, first route by capability: verify a basic OpenAI-compatible request for a unified SDK; verify the search call and sources used in the final answer for web search; verify tool parameters and multi-turn returns for agents; and verify rate limits, latency, tokens, and complete task cost for batch workloads. Model names only identify candidates and cannot replace real acceptance testing.

Conclusion first: how should Chinese LLM APIs be compared?

DeepSeek, Doubao, Qwen, Kimi, and Zhipu GLM should not be compared by brand name alone. First determine whether the team needs vendor-native features, OpenAI-compatible migration, a unified multi-model entry point, or web search and citations; then compare responses, tokens, latency, errors, and billing with the same model ID, prompt, and test window.

To reduce client changes, a team can include a unified service supporting OpenAI compatibility in a proof-of-concept, but seeing a model name does not mean that the model is usable. For AIFast, the public verification entry points are https://www.aifast.hk/v1/models andPlatform fact page; they confirm the current public entry points and verification method, while actual model access must still be confirmed with the team's own temporary key and real request.

For web search on Chinese LLMs, vendor-native request parameters and citation fields are not uniform. Use the official Doubao, Qwen, and Zhipu documentation as protocol references. Whether a unified gateway passes through these capabilities must be checked in the actual search calls, answer citations, and source URLs; compatibility alone does not prove identical web-search behavior.

Minimum reproducible record: test date, model ID, Base URL, HTTP status, request ID, time to first byte and total time, input and output tokens, SSE termination status, tool-call result, actual citation URLs, and billing changes. Do not record a complete API key or turn one successful request into a long-term stability conclusion.

If you are looking for a unified access method to DeepSeek API, Tongyi Qianwen API, Kimi API, Doubao API or Zhipu GLM, this article provides a repeatable OpenAI Compatible verification process; the specific model ID, open status and price are still subject to the real-time return of the API. You can also open it before accessingModel Quality CheckConfirm the target entrance before usingBase URL checkerTroubleshoot path splicing errors.

Check before you start

Names such as DeepSeek, Tongyi Qianwen, Kimi, Doubao, and GLM usually represent manufacturers or model series, and are not necessarily equal to the model ID that can be directly used by the API. The availability, price, context, and capabilities of different models will vary, and this article does not provide a fixed list.

Prepare the following information:

1. API Key created by the console.
2. Base URL https://www.aifast.hk/v1.
3. /v1/models Or the actual model ID displayed by the console.
4. Business must-have capabilities, such as streaming output, tool invocation, visual understanding, or retrieval.

Configuration steps

1. Get the real-time model list

curl https://www.aifast.hk/v1/models \
  -H "Authorization: Bearer $AIFAST_API_KEY"

Copy the model ID from the response. Don't pass family names like "DeepSeek" "Kimi" or "GLM" directly as request parameters unless the real-time API actually returns the exact same ID.

2. Complete minimal text call

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AIFAST_API_KEY"],
    base_url="https://www.aifast.hk/v1",
)

response = client.chat.completions.create(
    model="YOUR_MODEL_ID",
    messages=[{"role": "user", "content": "Explain your main abilities in three points。"}],
)

print(response.choices[0].message.content)

Only after the minimum text call is successful, continue to add streaming output, structured output, tools or multi-modal parameters.

3. Choose models based on capabilities rather than manufacturer name

The shared client code should also keep a platform-difference checklist: record model IDs, request protocols, streaming events, tool calls, web-search response fields, and citation URLs. This lets you reuse authentication and retry code without incorrectly presenting one platform's search or tool capability as supported by every Chinese LLM.

4. Preserve switchable configurations

Do not scatter model IDs through application code. Store the Base URL, API Key, and model ID in environment variables or a configuration service so staged rollouts and rollback remain manageable.

How to verify web-search capabilities

Web-search request parameters, response fields, and citation formats differ across model platforms. When a workflow depends on web results, retain the complete request, response ID, final answer, and actual cited URLs instead of checking only for a successful status.

If the product must display sources, verify that the final answer contains accessible citation URLs. A search-result list that is not used in the answer does not pass citation-display acceptance.

Next step

Continue reading How to choose text, image, video, and retrieval models and LLM API cost estimation and usage control, establish verifiable model selection and budgeting rules.