LLM API Guide

Quick start

curl https://air.cgu.edu.tw/cgullmapi/v1/models \
  -H "Authorization: Bearer YOUR_CGU_API_KEY"

Model catalog

These models are currently available to everyone. Put the model name from the tables into the model field of your request; /v1/models shows exactly which ones your account can use.

School compute pool (Spark) models | local model token quota

The university's own GPU servers. They support /v1/chat/completions and /v1/responses, and you can also use them for coding with opencode.

curl https://air.cgu.edu.tw/cgullmapi/v1/chat/completions \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-oss-120b",
    "messages": [
      {"role": "user", "content": "Introduce Chang Gung University in one sentence."}
    ]
  }'

import base64
from openai import OpenAI

client = OpenAI(
    api_key="YOUR_CGU_API_KEY",
    base_url="https://air.cgu.edu.tw/cgullmapi/v1",
)

with open("photo.png", "rb") as f:
    image = base64.b64encode(f.read()).decode()

response = client.chat.completions.create(
    model="gemma-4-26b-a4b-it",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "Describe this image."},
        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image}"}},
    ]}],
)
print(response.choices[0].message.content)

OpenAI models | OpenAI cost quota (USD)

Other, more expensive OpenAI models are currently locked; see the API Key page for the list.

Local models (Ollama) | local model token quota

If a model has been idle for a while, the first call has to load it first and takes longer. For images, use the school compute pool models gemma-4-26b-a4b-it or mistral-small-4-119b-2603 above.

Which API should I use?

Responses API

Recommended for new projects, Codex and agent workflows. The Responses API takes the user's input in input; with the SDK, the reply text is in response.output_text.

curl https://air.cgu.edu.tw/cgullmapi/v1/responses \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-luna",
    "input": "Introduce the CGU AI Center in one sentence.",
    "max_output_tokens": 128
  }'

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_CGU_API_KEY",
    base_url="https://air.cgu.edu.tw/cgullmapi/v1",
)

response = client.responses.create(
    model="gpt-5.6-luna",
    input="List three key points for testing an API.",
    max_output_tokens=160,
)
print(response.output_text)

Streaming

Streaming is good for showing text as it is generated, for long answers, or for cutting the perceived wait. When testing with curl, add -N so curl doesn't buffer the output.

curl -N https://air.cgu.edu.tw/cgullmapi/v1/responses \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-luna",
    "input": "Explain what a reverse proxy is for, in three points.",
    "stream": true,
    "max_output_tokens": 180
  }'

Chat Completions (compatibility)

Older SDKs, Open WebUI and tools that only support Chat Completions can use this route. It takes a messages array; the usual roles are system, user and assistant.

curl https://air.cgu.edu.tw/cgullmapi/v1/chat/completions \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-luna",
    "messages": [
      {"role": "system", "content": "You are a concise teaching assistant."},
      {"role": "user", "content": "Reply OK and confirm the API works."}
    ],
    "max_completion_tokens": 128
  }'

Embeddings

Embeddings turn text into numeric vectors. The vectors aren't meant to be read; databases and programs use them for similarity search, RAG document retrieval, duplicate detection, classification and recommendations.

curl https://air.cgu.edu.tw/cgullmapi/v1/embeddings \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-3-small",
    "input": [
      "CGU LLM Gateway test text",
      "This text will be turned into a vector"
    ]
  }'

Images: generation and editing

There are two image functions: /images/generations creates an image from a text prompt; /images/edits modifies, redraws, restyles or extends an uploaded image. chatgpt-image-latest, gpt-image-1, gpt-image-1-mini, gpt-image-1.5 and gpt-image-2 have all passed generation and editing tests; /v1/models shows which ones your account can use.

curl https://air.cgu.edu.tw/cgullmapi/v1/images/generations \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "chatgpt-image-latest",
    "prompt": "A clean blue square icon with CGU letters on a white background",
    "size": "1024x1024",
    "quality": "low",
    "n": 1
  }'

curl https://air.cgu.edu.tw/cgullmapi/v1/images/edits \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -F model=chatgpt-image-latest \
  -F image=@input.png \
  -F prompt="Keep the main composition and turn it into a clean blue-and-white tech-style icon" \
  -F size=1024x1024

Audio: speech-to-text, translation, text-to-speech

curl https://air.cgu.edu.tw/cgullmapi/v1/audio/transcriptions \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -F model=whisper-1 \
  -F language=en \
  -F response_format=json \
  -F file=@meeting.wav

curl https://air.cgu.edu.tw/cgullmapi/v1/audio/translations \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -F model=whisper-1 \
  -F response_format=json \
  -F file=@speech.wav

curl https://air.cgu.edu.tw/cgullmapi/v1/audio/speech \
  -H "Authorization: Bearer YOUR_CGU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o-mini-tts",
    "voice": "alloy",
    "input": "Today is a great day to finish your API integration.",
    "instructions": "Read in a natural, clear tone.",
    "response_format": "mp3"
  }' \
  --output speech.mp3

Realtime WebSocket

Realtime needs WebSocket and is good for low-latency voice, live conversation, live translation or live transcription. If you go through an external reverse proxy, make sure it forwards Upgrade: websocket, Connection: Upgrade and the related WebSocket headers.

const ws = new WebSocket(
  "wss://air.cgu.edu.tw/cgullmapi/v1/realtime?model=gpt-realtime-2",
  [],
);

ws.addEventListener("open", () => {
  ws.send(JSON.stringify({
    type: "session.update",
    session: { instructions: "Reply in English." }
  }));
});

ws.addEventListener("message", (event) => {
  console.log(JSON.parse(event.data));
});

Usage and billing

/v1/me/usage returns the local and OpenAI usage, token quota, OpenAI USD quota and remaining amount for the account bound to your API key. Check it before and after heavy testing to make sure your requests are counted correctly.

curl https://air.cgu.edu.tw/cgullmapi/v1/me/usage \
  -H "Authorization: Bearer YOUR_CGU_API_KEY"

Troubleshooting

Codex guide

1. Quit the Codex app completely

If you have used Codex before, sign out of your account in the Codex app first, then close the app.

2. Activate and copy your CGU API Key

3. Find the Codex settings folder

4. Edit config.toml

In the Codex settings folder, find or create config.toml and add the following:

model_provider = "custom"
model = "gpt-5.6-luna"
model_reasoning_effort = "medium"

[model_providers.custom]
name = "CGU API"
base_url = "https://air.cgu.edu.tw/cgullmapi/v1"
requires_openai_auth = true
wire_api = "responses"

Keeping your existing project settings

5. Restart the Codex app

6. Enter your CGU API Key when prompted

opencode guide (school compute pool models)

opencode is an open-source AI coding assistant that runs in your terminal: it reads code, edits files and runs commands for you. Connected to this service, it can use the school compute pool (Spark) models, which only count against your local model token quota and cost no USD.

1. Install opencode

2. Create opencode.json in your project folder

In your project folder (the one you want opencode to work on), create a file named opencode.json with the content below. It contains no key, so it's safe to commit to Git.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "cgu": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "CGU API",
      "options": {
        "baseURL": "https://air.cgu.edu.tw/cgullmapi/v1"
      },
      "models": {
        "gpt-oss-120b": { "name": "gpt-oss-120b (CGU compute pool)", "limit": { "context": 131072, "output": 32768 } },
        "gemma-4-26b-a4b-it": { "name": "gemma-4-26b-a4b-it (CGU compute pool)", "limit": { "context": 262144, "output": 32768 } },
        "mistral-small-4-119b-2603": { "name": "mistral-small-4-119b-2603 (CGU compute pool)", "limit": { "context": 262144, "output": 32768 } },
        "nemotron-3-super-120b-a12b": { "name": "nemotron-3-super-120b-a12b (CGU compute pool)", "limit": { "context": 262144, "output": 32768 } }
      }
    }
  },
  "model": "cgu/gpt-oss-120b"
}

To use this setup in every project on macOS / Linux, you can also put it in ~/.config/opencode/opencode.json.

3. Enter your CGU API Key (one time only)

4. Start using it

Just describe what you want, e.g. “Show me the structure of this project” or “How do I fix this error?”. The default model is gpt-oss-120b; type /models to switch. The first time in a project, you can type /init to have it write a project overview (AGENTS.md). If something goes wrong, /undo reverts the last step.