Pointing Cursor at DeepSeek, GLM or Qwen takes about two minutes and one setting: Cursor lets you override the OpenAI base URL, so any OpenAI-compatible endpoint becomes a model provider. This guide walks through the exact configuration, explains which Cursor features keep working and which do not, and does the cost arithmetic on what a full-time developer actually spends.
Cursor’s custom-model setting routes chat and inline edit requests to your endpoint instead of Cursor’s own backend. That gives you model choice and per-token billing you control. It is worth being precise about the trade-off up front, because a lot of writing on this topic is not:
If tab completion is the feature you value most, a custom endpoint supplements Cursor rather than replacing it. If chat and inline edit are where your time goes, this configuration covers the bulk of your usage.
Create a key in the console. Email or GitHub signup is sufficient — no Chinese phone number, no real-name verification, and billing is in USD. That access barrier is the main reason developers outside mainland China route these models through a gateway at all; see accessing Chinese AI models without a Chinese phone number.
Go to Settings → Models. You will find an OpenAI API key field and, below the model list, an option to override the base URL.
Base URL: https://aiwave.live/v1
API Key: sk-YOUR_KEYCursor sends requests to {base_url}/chat/completions, so the /v1 suffix is
required. Omitting it is the single most common configuration error and produces a 404 that looks like an
authentication problem.
Cursor’s model list is fixed unless you add entries. Use Add model and enter the model ID exactly as the endpoint serves it:
deepseek-v4-pro
deepseek-v4-flash
glm-5
qwen3-coder-480b-a35b-instruct
kimi-k2.5Model IDs are versioned and do change. Rather than trusting a list from a blog post, read the current one from the endpoint:
curl https://aiwave.live/v1/models \
-H "Authorization: Bearer sk-YOUR_KEY" | jq -r '.data[].id' | sortThe full catalogue is also on the model directory, and the endpoint is documented in the API reference.
Cursor’s verify button confirms the key and base URL. If it fails, check these in order:
/v1 in the base URL. Most common by a wide margin./v1/models output above; a
404 model_not_found means the ID is wrong, not the key.You can confirm the endpoint independently of Cursor with a plain curl, which isolates whether the problem is the API or the editor:
curl https://aiwave.live/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-YOUR_KEY" \
-d '{
"model": "deepseek-v4-pro",
"messages": [{"role": "user", "content": "Reply with OK"}]
}'Editor work splits into two very different workloads, and using one model for both wastes money or quality depending on which you pick.
Current rates for models commonly used in editors:
| Model ID | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
deepseek-v4-flash | $0.206 | $0.412 |
deepseek-v4-pro | $1.09 | $2.17 |
qwen3-coder-480b-a35b-instruct | $0.12 | $0.36 |
glm-4.7-flash | free | free |
glm-5 | $1.55 | $4.96 |
kimi-k2.5 | $0.66 | $3.30 |
Rates read from the AIWave pricing endpoint on 2026-07-26. Check live pricing before budgeting — providers revise rates.
Switch per request from Cursor’s model picker. Because the model is a configuration value, the right habit is to default to cheap and escalate deliberately, rather than paying premium rates for “add a docstring”.
Editor usage is bursty and easy to underestimate. A reasonable working assumption for a developer using chat and inline edit through the day is on the order of 20M input and 4M output tokens a month — context windows in an editor are large because the file, and often the surrounding files, go into every request.
The arithmetic is (input tokens ÷ 1,000,000) × input price + (output tokens ÷ 1,000,000) × output price. Run your own numbers against the table above rather than trusting a generic estimate; token consumption in an editor varies enormously with how much context you attach. The live pricing page always has current rates.
Two habits reduce the bill more than model choice does. Attach only the files that matter instead of whole directories, and start a new chat when you change task — a long thread resends its entire history on every message.
A .cursorrules file at the repository root gives the model persistent project context.
This matters more with a smaller model, which benefits from explicit constraints rather than inferring
conventions from surrounding code.
# .cursorrules
- TypeScript strict mode. No `any` without a comment explaining why.
- React function components only. No class components.
- Tests use Vitest, colocated as *.test.ts next to the source file.
- Prefer early returns over nested conditionals.
- Do not add dependencies without asking; this repo pins versions deliberately.
- Error handling: throw typed errors from lib/errors.ts, never bare strings.Keep it short and specific. Long rule files consume context on every request, which costs money and crowds out the code you actually want the model to read.
Routing through a custom endpoint means your code snippets go to that endpoint rather than to Cursor’s backend. That is a change in who processes your source, not a removal of processing. Enable Cursor’s privacy mode if you want to keep code out of Cursor’s own storage, and treat the API provider as you would any third-party service in your supply chain. For regulated codebases, confirm your own compliance requirements before enabling any AI editor, custom endpoint or not.
Almost always a stale setting: another provider still enabled, or a model name with a typo. Disable unused providers and re-add the model.
Expected. Tab completion runs on Cursor’s proprietary model and is not routed through custom endpoints.
Agent flows lean heavily on tool calling, which is an upstream model capability. If a model handles tools inconsistently, agent mode will be unreliable with it. Check the model page, and see function calling with Chinese AI models for the protocol-level detail.
The upstream provider is overloaded. Switch to another model in the picker — that is the practical benefit of having several configured. Error codes covers which failures are transient.
The same base-URL override works in most AI editor tooling, because they all speak the OpenAI wire format: Cline and Continue in VS Code, Aider on the command line, Zed’s assistant, and JetBrains plugins that accept a custom endpoint. Related walkthroughs: DeepSeek V4 in Cursor, Continue in VS Code, Kilo Code in VS Code and OpenCode on the CLI.
Editor prompting is a different discipline from chat prompting. The model already has code in front of it, so the value you add is constraint, not description. Three patterns account for most of the difference between useful and frustrating results.
“Make this faster” invites the model to rewrite whatever it likes. “Reduce the allocations in this loop without changing the public signature or the error semantics” produces a diff you can actually review. The more precisely you fence the change, the smaller and safer the result.
Pasting the stack trace, the failing assertion or the actual versus expected value is worth more than any amount of prose. Models are good at pattern-matching an error to a cause and poor at guessing what went wrong from a description of intent.
Requesting a minimal patch rather than a rewritten file keeps output tokens down, keeps review tractable, and dramatically reduces the chance of the model silently dropping code it did not understand. This matters more with cheaper models, which are likelier to omit sections when regenerating long files.
Bad: "Refactor this file."
Good: "Extract the retry logic from lines 40-72 into a helper in lib/retry.ts.
Keep the existing exponential backoff behaviour. Show only the diff."Cursor sends what you attach. That is both the feature and the cost driver — every attached file is input tokens on every message in the thread, not just the first.
.cursorignore to keep build output, lockfiles, fixtures and
vendored code out of indexing. These inflate context with material that never helps.# .cursorignore
node_modules/
dist/
build/
coverage/
*.lock
**/__snapshots__/
**/fixtures/**A useful mental model: you are paying for the model to read. Give it the chapter, not the library.
Configuring several models is only useful if you actually switch between them. The workflow that works in practice is deliberately boring:
The second step is where the saving is. Most teams that adopt a premium model everywhere are paying premium rates for work a cheap model handles indistinguishably. The same principle applied programmatically is covered in building a multi-model router, and the cost side in AI API cost optimisation.
The failure mode of a fast editor assistant is not bad code that obviously breaks. It is plausible code that passes review because it looks like the surrounding style.
None of this is specific to Chinese models — it applies to every AI coding assistant. It is worth stating because the speed of a cheap fast model makes it easy to accept changes faster than you can genuinely evaluate them.
If you are moving a team rather than yourself, a few things smooth it considerably.
.cursorrules and .cursorignore to the repository so
everyone gets the same conventions and the same context discipline.Being honest about this matters more than selling the configuration.
Where it clearly wins: you want a specific model Cursor does not offer, you want to control per-token spend, or you need models that would otherwise require a Chinese phone number and a mainland payment method to access.
Yes. Set the OpenAI base URL to an OpenAI-compatible endpoint, add your key, and add the DeepSeek model IDs manually. Chat and inline edit then run on DeepSeek.
No. Cursor’s tab completion uses its own proprietary model and is not routed through custom endpoints. Chat and inline edit are.
Custom API keys work without one, though feature availability varies by Cursor version. You pay the API provider per token either way.
Try a code-specialised model such as the Qwen coder variant and a strong general model such as DeepSeek V4 Pro on your own repository. Editor performance depends heavily on your language and codebase, so a week of real use tells you more than any leaderboard.
Every claim above traces back to one of these. Pricing figures come from the AIWave pricing endpoint on 2026-07-26; capability claims come from the vendors’ own documentation.