# Chinese Models

QCode.cc sells more than Claude / GPT / Gemini. The same `cr_` key also reaches Zhipu GLM, Moonshot Kimi, DeepSeek and Qwen. This page only answers "how do requests land on these ids" — no style comparisons, no marketing.

The daily default should still be the Claude 5 family (see [Model Selection Guide](/docs/usage/model-selection)). The Chinese families shine at: Chinese office and long-form text, cheaper bulk volume, and as a control against Claude. Never put a name into a client that does not appear on [qcode.cc/models](https://qcode.cc/models).

## What is on sale now

The ids below appeared on 2026-09-18 both on [qcode.cc/models](https://qcode.cc/models) and on the public endpoint `GET https://api.qcode.cc/api/v1/models` (cross-checked; live unit prices and context sizes are whatever [qcode.cc/models](https://qcode.cc/models) shows — this page deliberately does not copy numbers).

| Family | Flagship id | One tier lighter |
|------|---------|----------|
| Zhipu GLM | `glm-5.3` | `glm-5.3-flash` |
| Moonshot Kimi | `kimi-k3` | — (previous `kimi-k2.6` has been retired) |
| DeepSeek | `deepseek-v4-pro` | `deepseek-v4-flash` · `deepseek-v4.1-flash` |
| Qwen | `qwen3.8-max` | `qwen3.8-flash` · `qwen3.7-plus` |

Names you may still see in old examples but which are no longer sold: `qwen3.7-max`, `kimi-k2.6`, `glm-5.1` — gone from both live lists; entering them in a client will fail, switch to the ids above.

**No grok.** The live list: [qcode.cc/models](https://qcode.cc/models) or `GET https://api.qcode.cc/v1/models` with your key.

## Which protocol

The protocol is decided by the **path**; the key does not care about families. Source of truth: [Endpoints and API paths](/docs/getting-started/endpoints-and-api-paths).

| Your tool | Fill in | Model field takes |
|------------|--------|----------------|
| OpenAI SDK / LangChain / Cline "OpenAI Compatible" / [WorkBuddy](/docs/ide/workbuddy) | `https://api.qcode.cc/openai/v1` (mainland China: prefer `https://asia.qcode.cc/openai/v1`) | ids from the table above, character for character |
| Claude Code (already pointed at QCode per [Environment](/docs/getting-started/environment)) | no need to change `ANTHROPIC_BASE_URL` | in-session `/model glm-5.3` |
| Hand-written HTTP | `POST /openai/v1/chat/completions` | `"model"` in the JSON |

Do not put Chinese-family ids behind the Gemini `/gemini` prefix, and do not point OpenAI-compatible tools at `https://api.qcode.cc/api` (that is the Anthropic Messages prefix).

Non-Claude models may lose native features like Extended Thinking inside Claude Code — [Model Selection Guide](/docs/usage/model-selection) documents the same caveat for GPT. When in doubt, treat them as a supplementary tier, not a replacement.

## curl self-test

Same interpretation as the endpoints doc: `400` = path and key both fine (missing body expected), `401` = key invalid, `404` = wrong prefix.

```bash
KEY="cr_your-QCode-key"

# path check only
curl -s -o /dev/null -w '%{http_code}\n' -X POST https://api.qcode.cc/openai/v1/chat/completions \
  -H "Authorization: Bearer $KEY"

# end-to-end with one id on sale (replace glm-5.3 with the id you want)
curl -sS https://api.qcode.cc/openai/v1/chat/completions \
  -H "Authorization: Bearer $KEY" \
  -H "content-type: application/json" \
  -d '{"model":"glm-5.3","messages":[{"role":"user","content":"ping"}]}'
```

From mainland China swap the host to `asia.qcode.cc` and test again. One key works on all three domains.

Reference OpenAI Python client (set `base_url` to `/openai/v1`, **no** trailing slash):

```python
from openai import OpenAI

client = OpenAI(
    api_key="cr_your-QCode-key",
    base_url="https://api.qcode.cc/openai/v1",
)
resp = client.chat.completions.create(
    model="glm-5.3",
    messages=[{"role": "user", "content": "ping"}],
)
```

## Pairing with the Claude 5 family

| What you are doing | Start with | When the Chinese models show up |
|------------|------|----------------|
| Day-to-day code edits | `claude-sonnet-5` | run a control pass, or use `deepseek-v4-pro` / `glm-5.3` on a budget |
| Hard reasoning / big refactors | `claude-opus-5` | no forced switch; try `kimi-k3` to save |
| Long Chinese text, notes, office | any flagship | `kimi-k3`, `glm-5.3`, `qwen3.8-max` are common |
| Bulk grunt work | `claude-haiku-4-5` | `deepseek-v4-flash` / `qwen3.8-flash` cost less |

The 4.x generation (`claude-sonnet-4-6`, `claude-opus-4-8`) is **still on sale**, just no longer the default. Prices and windows: [qcode.cc/models](https://qcode.cc/models).

Same key in an office desktop client: see [WorkBuddy Integration](/docs/ide/workbuddy). Switching Claude Code / Codex: [CC Switch Setup](/docs/ide/cc-switch).

## FAQ

### Filled the id but got 404 / an empty list

The id must look like `glm-5.3` in the table, not "GLM-5.3" or a display name. OpenAI-compatible mode often lists nothing — type the id manually.

### 401

The key starts with `cr_`, no stray spaces. Verify in [qcode.cc/dashboard](https://qcode.cc/dashboard).

### Can I call GLM over Anthropic's `/api/v1/messages`?

The gateway picks the protocol by path. `/openai/v1/chat/completions` is the documented path for OpenAI-compatible clients. With Claude Code already pointed at QCode, just `/model <id>` — no hand-built Messages calls for the Chinese tier.

### How about WorkBuddy / Cline

WorkBuddy: provider Custom, endpoint `https://api.qcode.cc/openai/v1`, model name = an id from the table; step-by-step in [WorkBuddy Integration](/docs/ide/workbuddy). Cline uses the same Base URL and Model ID under "OpenAI Compatible".

## Next steps

- [Endpoints and API paths](/docs/getting-started/endpoints-and-api-paths)
- [Model Selection Guide](/docs/usage/model-selection)
- [Billing](/docs/reference/billing)
- [WorkBuddy Integration](/docs/ide/workbuddy)
- Live ids: [qcode.cc/models](https://qcode.cc/models)