Chinese LLM API free trial: how it works and what people ask

About the trial

What do I get for free?

Every new account receives $0.50 of credit. It is valid for seven days from sign-up, and you do not need to enter payment details to claim it. The credit applies to the same endpoint and the same model that paid accounts use, so what you test is what you would run in production.

How far does $0.50 go?

It depends on your prompts, so here is an illustration with assumptions spelled out. Suppose each request sends 1,000 input tokens and receives 500 output tokens. At $0.25 per million input tokens and $1.00 per million output tokens, one request costs $0.00025 plus $0.0005, which is $0.00075. Then $0.50 covers roughly 660 such requests. Shorter exchanges stretch further, and long-context calls use the credit faster.

What happens when the trial ends?

If the seven days pass or the credit is spent first, requests return HTTP 402 with the code no_credit. Nothing is deleted. Top up prepaid balance and the same key works again. There is no subscription, and topped-up balance never expires.

Do I need to add payment details to start?

No. Registration takes an email and a password only. The key is displayed right after you sign up at the get-your-key page.

A practical plan for the seven days

Seven days is plenty if you decide in advance what you want to learn. Treat the trial like a small evaluation project rather than a place to poke around. Here is a sequence that works well for most teams:

1. Day 1, smoke test. Confirm the key, the model list and one short completion. The commands below do all three.
2. Days 2 to 3, your real prompts. Run twenty to fifty representative inputs from your own product through the API. Save the outputs next to the inputs so a colleague can review them without rerunning anything.
3. Day 4, failure paths. Deliberately trigger a 400 with an oversized request, a 401 with a wrong key, and test your 429 and 503 handling with a small burst. It is better to see these now than in production.
4. Days 5 to 6, cost model. Multiply your logged token counts by the published rates and project a monthly figure under your own traffic assumptions.
5. Day 7, decision. Compare quality, cost and limits against your requirements, then decide whether to top up.

Copyexport API_KEY="paste-your-key-here"

# 1) confirm the key works and see the model id
curl -s https://api.chinesellmapi.com/v1/models \
  -H "Authorization: Bearer $API_KEY"

# 2) one short completion; read the usage block in the reply
curl -s https://api.chinesellmapi.com/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"uncensored","messages":[{"role":"user","content":"用一句话介绍杭州的春天。"}],"max_tokens":120}'

The second call returns a usage object with prompt and completion token counts. Log those from the first day onward; they are the raw material for every cost estimate you will make later.

Keys and getting started

How many keys can I have?

One per account. If you need a new one, regenerate it from your account. The old key stops working immediately, so update your deployments straight away, or you will see 401 errors from anything still using the previous value.

Where do I put the key?

In an environment variable on your server, named for example API_KEY, and never in browser code or a public repository. Requests carry it as Authorization: Bearer <key>.

What is the smallest request that proves it works?

List the models with GET /v1/models, then send a short chat completion to /v1/chat/completions with model uncensored and a low max_tokens. The usage field in the response shows exactly what the call cost in tokens.

Can I use the OpenAI SDK?

Yes. The API is compatible with OpenAI chat completions. Point the base URL at https://api.chinesellmapi.com/v1, supply your key, and set the model to uncensored. A full walkthrough, including prompts and script control, is in the quickstart.

Limits and capabilities

How long can a prompt and reply be?

The context window is 100,000 tokens, counting prompt and completion together. The max_tokens setting defaults to 2,048 and can be raised to 16,000 per request. A request body can be up to 8 MB. If prompt plus max_tokens exceeds the window, the API returns 400.

How many requests can I send?

Each key is limited to 300 requests per minute. Beyond that you receive 429. Spread batch jobs evenly and add a backoff; the translation guide includes a paced batch script.

Does it support streaming and function calling?

Yes to both. Set stream: true for server-sent events, with a final chunk carrying usage. Tools are accepted in OpenAI format, and sampling fields such as temperature, top_p and stop are passed through.

What is not available?

Only text is offered: no embeddings, images, audio, video or fine-tuning. There is a single model. If your project needs one of the missing features, plan to pair this API with another tool for that part.

Errors you may meet

Which status codes should I expect?

Errors return JSON shaped like {"error":{"code":...,"message":...}}. A 400 means a malformed request, 401 an invalid or missing key, 402 no_credit, 403 content_blocked, 404 an unknown endpoint, 429 the rate limit and 503 upstream_busy.

Which errors are worth retrying?

Only 429 and 503. Wait, add random jitter, and retry a limited number of times; a few seconds is enough for 503. Retrying 400, 401, 402 or 403 will not change the outcome.

I topped up but still see 402, what now?

Check that you are using the current key, since regenerating a key replaces the old one. Then confirm the request is going to the right base URL. If everything matches, send a minimal request and read the error message in the response body.

Making the trial credit last

The credit is small by design, so a few habits stretch it. Keep max_tokens low while you are testing, since output tokens cost four times as much as input tokens: $1.00 per million versus $0.25 per million. Test with short, representative prompts first and move to your longest inputs only once the short ones behave. Avoid loops that call the API without a stop condition, and cap the number of retries in your own code.

Worked example with stated assumptions: an evaluation set of 50 prompts, each about 600 input tokens, with replies capped at 300 output tokens. One pass costs 30,000 input tokens plus 15,000 output tokens, which is $0.0075 plus $0.015, or $0.0225. You could run that whole set around 22 times before the credit is gone, which is enough to compare several prompt variants properly.

If you pass a very long prompt, remember that the whole prompt is billed on every call. A 40,000-token document sent twenty times costs the same as twenty separate large requests, so cache or summarize shared context where you can. Current rates are listed on the pricing page, and parameter details are in the docs.

Content rules and privacy

What content is allowed?

Lawful adult content, fiction and controversial topics are not refused as such. The service is for adults aged 18 or over. For long-form fiction and character chat, the fiction guide covers prompt structure and cost.

What is always blocked?

Sexual content involving minors is blocked in every case and returns a 403, including fiction and roleplay. The block cannot be changed by prompt wording, so do not retry it.

Are my prompts used for training?

No, prompts are not used for training.