Quinn Qwen3.5 Omni Plus Realtime

GET 
/v1/realtime

OpenAI Realtime-style WebSocket for qwen3.5-omni-plus-realtime. Video is sent as JPEG frames, not a raw video stream.

Snapshot alias qwen3.5-omni-plus-realtime-2026-03-15 has the same capability.

Use a platform-issued API token. Do not send OpenAI Sec-WebSocket-Protocol (openai-insecure-api-key.*); this gateway and the upstream both use Bearer.

Connect

This is a WebSocket endpoint (wss://), not an HTTP GET. Do not call it with a normal HTTPS curl.

wss://open-api.fancyai.com/v1/realtime?model=qwen3.5-omni-plus-realtime

Authorization: Bearer YOUR_API_KEY

curl (needs curl 8+ with WebSocket support):

curl --no-buffer \
  -H "Authorization: Bearer $API_KEY" \
  "wss://open-api.fancyai.com/v1/realtime?model=qwen3.5-omni-plus-realtime"

After the handshake, the server sends session.created.

Python:

import os
import json
import websocket

WS_URL = "wss://open-api.fancyai.com/v1/realtime?model=qwen3.5-omni-plus-realtime"
API_KEY = os.environ["API_KEY"]

def on_open(ws):
    print("connected")

def on_message(ws, message):
    print(json.dumps(json.loads(message), ensure_ascii=False, indent=2))

ws = websocket.WebSocketApp(
    WS_URL,
    header=["Authorization: Bearer " + API_KEY],
    on_open=on_open,
    on_message=on_message,
)
ws.run_forever()

Request

Responses

101

Switching Protocols — WebSocket established

400

Invalid request parameters

401

Unauthorized

429

Rate limit exceeded

Configure the session

Send session.update after connect:

{
  "event_id": "event_001",
  "type": "session.update",
  "session": {
    "modalities": ["text", "audio"],
    "voice": "Tina",
    "instructions": "You are a voice assistant. Answer briefly in spoken language.",
    "audio": {
      "input": { "format": { "type": "pcm", "sample_rate": 16000 } },
      "output": { "format": { "type": "pcm", "sample_rate": 24000 } }
    },
    "turn_detection": {
      "type": "semantic_vad",
      "threshold": 0.5,
      "silence_duration_ms": 800
    }
  }
}

Field	Notes
modalities	["text"] text only; ["text","audio"] text + speech
voice	Output voice, default Tina
audio.input/output.format	Qwen 3.5 Realtime only. Input default pcm/16000, output default pcm/24000. Sample rate: 8000 / 16000 / 24000 / 48000
input_audio_format / output_audio_format	Legacy fields, still accepted
turn_detection	semantic_vad (recommended) or server_vad. null = manual mode
enable_search	Web search, off by default. Cannot be used together with tools
tools	Function tools. Cannot be used together with enable_search

Server replies with session.updated.

End the session with session.finish, or close the WebSocket. Leaving it open keeps accumulating context and billing.

Send audio

Base64-encode PCM/WAV and send:

{
  "type": "input_audio_buffer.append",
  "audio": "<base64>"
}

About 100ms per packet is recommended.

VAD mode (semantic_vad / server_vad): the server auto-commits and replies after it detects you finished speaking.

Manual mode (turn_detection is null): after audio, submit and trigger a reply:

{ "type": "input_audio_buffer.commit" }

{ "type": "response.create" }

Cancel the current reply:

{ "type": "response.cancel" }

Send images (video frames)

The model does not accept raw video. Extract JPEG frames and send Base64:

{
  "type": "input_image_buffer.append",
  "image": "<base64 jpeg>"
}

Constraints:

• Send input_audio_buffer.append at least once first
• JPG/JPEG; 480P or 720P recommended, max 1080P
• Base64 payload ≤ 256KB per frame; raw size ≤ 190KB recommended
• Live camera: 1 fps (max 2 fps)
• In VAD mode, finish the current frame before input_audio_buffer.speech_stopped

Audio is the timeline; images are inserted at send time. Camera can be turned on or off at any time.

Text-only input

{
  "type": "conversation.item.create",
  "item": {
    "type": "message",
    "role": "user",
    "content": [{ "type": "input_text", "text": "What's the weather outside?" }]
  }
}

In manual mode, follow with response.create.

Receive replies

Text only:

Event	Meaning
response.text.delta	Incremental text
response.text.done	Full text

Text + audio:

Event	Meaning
response.audio_transcript.delta	Incremental transcript of speech
response.audio_transcript.done	Full transcript
response.audio.delta	Base64 audio chunk; decode and play
response.audio.done	This audio segment ended
response.done	Turn finished, includes usage

Other events:

Event	Meaning
session.created	Connection ready
session.updated	Session config applied
input_audio_buffer.speech_started	User started speaking (you may cancel the current reply)
input_audio_buffer.speech_stopped	User finished speaking
conversation.item.input_audio_transcription.completed	User speech transcribed (if input transcription is on)
error	Error

Billing

Billed from usage on each response.done, not by connection duration. In multi-turn chats, audio / image / text still in the window is counted again as this turn's input.

When the reply includes speech, output text is not billed separately (matches gateway settlement).

Singapore list price converted at CNY/USD 6.8:

Item	plus-realtime
Input: text / image / video frame	$2.315 / M
Input: audio	$18.184 / M
Output: text only	$13.666 / M
Output: text + audio (output text not billed)	$68.329 / M

Audio conversion (round up to 1 second):

• Input: seconds × 7
• Output: seconds × 12.5

Images / video frames: about 1 token per 32×32 pixels, min 4 / max 1280 per image.

Web search is billed extra and is off by default. tools and enable_search cannot both be on.

Limits

	plus-realtime
Max audio turns	100
Max video (frame) turns	50
Max audio duration	600 s
Max video (frame) duration	240 s
Single session	120 minutes

The server disconnects on timeout. Older history is dropped.

Choose a model

Prefer plus for stronger understanding and longer context. For lower latency, use qwen3.5-omni-flash-realtime. This API does not provide WebRTC / AOQ.

Official capability notes: Qwen-Omni realtime. Clients must connect only to this platform /v1/realtime.