Stream audio and text over a real-time connection.
Qwen3.8-Omni-Flash-Realtime
New Launched on IteraCompute
qwen/qwen3.8-omni-flash-realtime- Tool calling
UnsupportedTool calling · Unsupported
- Reasoning
UnsupportedReasoning · Unsupported
- Vision
SupportedVision · Supported
- Streaming
SupportedStreaming · Supported
- JSON mode
UnsupportedJSON mode · Unsupported
USD per million tokens
- Text / image / video input
- $0.23
- Audio input
- $0.93
- Text output
- $0.70
- Audio output
- $1.87
- Precision
- Official precision
- Price checked
Model availability and prices
Overview
Send audio over a WebSocket and optionally append image frames before committing the audio turn. Receive audio chunks and a transcript, then read usage by modality in response.done.
- Version
- qwen3.8-omni-flash-realtime
- Input / output
- text, audio, image, video / text, audio
- Endpoint
- /v1/realtime
API protocols and integration
Authorization: Bearer ITERACOMPUTE_API_KEY
WebSocket /v1/realtime
Connect using WebSocket. Read usage from response.done.
LLMs.txt
# qwen/qwen3.8-omni-flash-realtime
Revision: 60543f71d943c248156389ae7199301bfebbd3586618b90a2c030cdec810f601
Authorization: Bearer ITERACOMPUTE_API_KEY
WebSocket wss://api.iteracompute.com/v1/realtime
- modalities · array · Requested response modalities.
- instructions · string · Instructions for this session.
- input_audio_format · enum · 16 kHz mono PCM16, little endian. · ["pcm"]
- output_audio_format · enum · 24 kHz mono PCM16. · ["pcm"]
- turn_detection · object · Use null for manual commit or configure voice activity detection.
- voice · string · A voice supported for this exact model.
Install websocket-client and set ITERACOMPUTE_API_KEY. Provide input.pcm as raw mono 16 kHz little-endian PCM16. The reply is saved as a playable 24 kHz WAV file. For visual input, send an audio chunk first, then input_image_buffer.append with a Base64 JPEG in the image field.
import base64, json, os, wave
from pathlib import Path
from urllib.parse import urlencode
from contextlib import closing
from websocket import create_connection
headers = {"Authorization": "Bearer " + os.environ["ITERACOMPUTE_API_KEY"]}
url = "wss://api.iteracompute.com/v1/realtime?" + urlencode({"model": "qwen/qwen3.8-omni-flash-realtime"})
audio = Path("input.pcm").read_bytes()
with closing(create_connection(url, header=headers, timeout=60)) as ws:
def receive():
event = json.loads(ws.recv())
if event["type"] == "error":
raise RuntimeError(event.get("error"))
return event
def wait_for(kind):
while receive()["type"] != kind:
pass
wait_for("session.created")
ws.send(json.dumps({"type": "session.update", "session": {"modalities": ["text", "audio"], "turn_detection": None}}))
wait_for("session.updated")
for offset in range(0, len(audio), 3200):
chunk = base64.b64encode(audio[offset:offset+3200]).decode("ascii")
ws.send(json.dumps({"type": "input_audio_buffer.append", "audio": chunk}))
ws.send(json.dumps({"type": "input_audio_buffer.commit"}))
wait_for("input_audio_buffer.committed")
ws.send(json.dumps({"type": "response.create"}))
with wave.open("reply.wav", "wb") as out:
out.setparams((1, 2, 24000, 0, "NONE", "PCM"))
while True:
event = receive()
if event["type"] == "response.audio.delta":
out.writeframesraw(base64.b64decode(event["delta"]))
elif event["type"] == "response.done":
if event["response"]["status"] != "completed":
raise RuntimeError(event["response"]["status"])
break
https://www.alibabacloud.com/help/en/model-studio/realtimeAPI parameters
- Connect using WebSocket. Read usage from response.done.
These parameters and ranges describe the current route. Account and context limits also apply. Parameter acceptance does not establish model capability.
- Tool calling
- Unsupported
- Reasoning
- Unsupported
- Vision
- Supported
- Streaming
- Supported
- JSON mode
- Unsupported
Parameters and values
Click a parameter to copy its name; scroll the table horizontally to see all columns. Usage notes identify protocol differences; identical names do not guarantee cross-endpoint support.
| Parameter | Type | Acceptance | Range | Usage / protocol |
|---|---|---|---|---|
| string | Accepted | — | A voice supported for this exact model. | |
| array | Accepted | — | Requested response modalities. | |
| string | Accepted | — | Instructions for this session. | |
| object | Accepted | — | Use null for manual commit or configure voice activity detection. | |
| enum | Accepted | ["pcm"] | 16 kHz mono PCM16, little endian. | |
| enum | Accepted | ["pcm"] | 24 kHz mono PCM16. |
Current configuration example
WebSocket /v1/realtime
{
"type": "session.update",
"session": {
"modalities": [
"text",
"audio"
],
"turn_detection": null
}
}Parameter revision:60543f71d943
API
Install websocket-client and set ITERACOMPUTE_API_KEY. Provide input.pcm as raw mono 16 kHz little-endian PCM16. The reply is saved as a playable 24 kHz WAV file. For visual input, send an audio chunk first, then input_image_buffer.append with a Base64 JPEG in the image field.
import base64, json, os, wave
from pathlib import Path
from urllib.parse import urlencode
from contextlib import closing
from websocket import create_connection
headers = {"Authorization": "Bearer " + os.environ["ITERACOMPUTE_API_KEY"]}
url = "wss://api.iteracompute.com/v1/realtime?" + urlencode({"model": "qwen/qwen3.8-omni-flash-realtime"})
audio = Path("input.pcm").read_bytes()
with closing(create_connection(url, header=headers, timeout=60)) as ws:
def receive():
event = json.loads(ws.recv())
if event["type"] == "error":
raise RuntimeError(event.get("error"))
return event
def wait_for(kind):
while receive()["type"] != kind:
pass
wait_for("session.created")
ws.send(json.dumps({"type": "session.update", "session": {"modalities": ["text", "audio"], "turn_detection": None}}))
wait_for("session.updated")
for offset in range(0, len(audio), 3200):
chunk = base64.b64encode(audio[offset:offset+3200]).decode("ascii")
ws.send(json.dumps({"type": "input_audio_buffer.append", "audio": chunk}))
ws.send(json.dumps({"type": "input_audio_buffer.commit"}))
wait_for("input_audio_buffer.committed")
ws.send(json.dumps({"type": "response.create"}))
with wave.open("reply.wav", "wb") as out:
out.setparams((1, 2, 24000, 0, "NONE", "PCM"))
while True:
event = receive()
if event["type"] == "response.audio.delta":
out.writeframesraw(base64.b64decode(event["delta"]))
elif event["type"] == "response.done":
if event["response"]["status"] != "completed":
raise RuntimeError(event["response"]["status"])
breakPricing
USD per million tokens
- Model ID
qwen/qwen3.8-omni-flash-realtime- Verification scope and sources
View verification scope and sources
2026-11-21 · Singapore WebSocket tests verified audio, red and blue image recognition, independent billing for sequential turns, and estimated settlement after a confirmed cancellation when supplier usage was missing.
Usage limits
- Audio and its accompanying text are billed separately.
Use cases
Integration tools and guides
FAQ
How do I get started?
How is usage billed?
New arrivals
Explore the models selected for New Arrivals, with API pricing, supported capabilities and integration guides.
Qwen-Audio-3.1-Realtime-Plus
qwen/qwen-audio-3.1-realtime-plus
Stream audio and text over a real-time connection.
Launched on IteraCompute
Qwen3.7-Plus
qwen/qwen3.7-plus
Qwen3.7-Plus returns text from text and visual input.
Launched on IteraCompute
- Input
- $0.384
- Output
- $1.536
USD / 1M tokens
Wan 3.0 Prime
wan/wan3.0-video-prime
Wan 3.0 Prime generates video from a prompt and optional reference media.
Launched on IteraCompute