Alibaba’s Qwen staff has launched Qwen3.8-Omni-Flash. They referred to as it its first omni-modal mannequin constructed round agentic capabilities. It accepts textual content, photos, audio, and video, and it returns textual content. Audio-video understanding, reasoning, and power use sit inside one mannequin. The said workflow is straightforward: perceive the content material, plan the duty, execute with instruments, ship the outcome.
Is it deployable? Sure, as a hosted API right now. It’s dwell on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio. No open weights had been introduced at launch, so self-hosting isn’t an choice.
What’s Qwen3.8-Omni-Flash
The Qwen3.8-Omni-Flash is constructed on the Qwen3.8-Flash-Next structure. That base mannequin shipped with open weights in August 2026.
The context window is 1M tokens. QwenCloud lists 991K max enter and 131K max output. Max reasoning size is 262K tokens.
Output is textual content solely. The Mannequin Studio docs level builders to Qwen3.5-Omni once they want generated speech. Considering is on by default, with reasoning_effort set to xhigh. Setting it to none disables considering.
The API follows each the DashScope and OpenAI protocols. It really works with Chat Completions and the Responses API. Perform calling, net search, structured outputs, context caching, and batch calls are supported.
Agentic Notion for Lengthy Video
Most video fashions learn an extended file from begin to end. That holds even when the reply sits in 3 minutes of footage.
Qwen analysis staff describes a distinct path. The agent begins from the query. It decides what to look at and listen to. It then gathers proof over a number of coarse-to-fine rounds. Compute and tokens go to the segments that matter.
The analysis staff stories the outcome on OmniVideoBench. Accuracy rises from 63.4 to 67.8. Token use drops from 145,736 to 79,117. That’s about 45.7% fewer tokens.
Reported Benchmarks
All figures right here come from Qwen. Unbiased outcomes weren’t out there at publication.
- Throughout 29 evaluations, the common rating improves greater than 25% over Qwen3.5-Omni-Plus.
- WildClawBench-MM improves by 36.5 factors. AgenticVBench improves by 22.3 factors.
- UniClawBench reaches 69.6.
- LongAudioSpan positive aspects 8.3 factors. OmniVideoBench positive aspects 9.6 factors.
- OmniCap-IF CSR and ISR enhance by 8.5 and 14.1 factors.
The analysis staff states that audio-visual efficiency is near Gemini 3.8 Flash. It additionally claims general audio efficiency above Gemini 3.8 Flash. The X post summarizes the agent positive aspects as +19.5 factors on common throughout WildClawBench-MM and UniClawBench.
Pricing and Enter Limits
QwenCloud lists $0.15 per 1M enter tokens and $0.47 per 1M output tokens. Implicit cache hits value $0.016 per 1M tokens.
The analysis staff stories massive value cuts in opposition to Qwen3.5-Omni-Plus. Audio enter prices over 98% much less per hour. Audio-visual enter prices over 93% much less per hour. The X submit places the video enter discount at about 89%.
Key limits from the Model Studio docs:
- Video information as much as 2 hours and a pair of GB by URL.
- Audio information as much as 3 hours.
- Audio enter in 113 languages and dialects.
- Steady outcomes with video sampled at as much as 15 fps.
- Two-channel stereo and four-channel FOA spatial audio by way of
use_multichannel. - Availability in 6 areas: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia.
Calling it takes just a few strains with the OpenAI SDK:
import os
from openai import OpenAI
shopper = OpenAI(
api_key=os.environ["DASHSCOPE_API_KEY"],
base_url=os.environ["DASHSCOPE_BASE_URL"],
)
completion = shopper.chat.completions.create(
mannequin="qwen3.8-omni-flash",
messages=[{"role": "user", "content": [
{"type": "video_url", "video_url": {"url": os.environ["VIDEO_URL"]}},
{"type": "text", "text": "List the key moments with timestamps."},
]}],
modalities=["text"],
stream=True,
)
for chunk in completion:
if chunk.decisions and chunk.decisions[0].delta.content material:
print(chunk.decisions[0].delta.content material, finish="")
The mannequin returns textual content, so instruments do the media work. Qwen staff is open-sourcing 2 tasks to assist that.
Qwen-MM-Plugins is dwell underneath Apache-2.0. Its tagline is ‘Make any agent harness multimodal-native.’ Every functionality installs as a Ability plus an elective MCP server. The guided installer helps Claude Code, CodeBuddy, Codex, Qoder, OpenClaw, Qwen Code, and Gemini CLI.
The Omni capabilities map to the launch demos:
omni-memorybuilds an audio-visual reminiscence of an extended video.omni-video2noteconverts a tutorial video into an illustrated PDF.omni-chatcutcovers Music-to-MV, film commentary, and speaker-preserving video translation.
A core plugin lets the primary mannequin learn native photos and video frames natively. The README notes one present hole. Most harnesses can’t feed audio to the primary mannequin natively but. Audio is routed by way of the API for now.
Interactive Explainer
/* 2 notion */
var tl=$(‘#tl’),cells=[];for(var i=0;i=plan.size){finalState();return}
var p=plan[n];cells[p[0]].className=p[1];$(‘#stg’).textContent=p[2];
n++;var f=n/plan.size;setK(F.tok*f,F.acc*f);
setTimeout(tick,step);
})();
}
$$(‘#mode button’).forEach(operate(b){b.addEventListener(‘click on’,operate(){
$$(‘#mode button’).forEach(operate(x){x.classList.toggle(‘on’,x===b)});mode=b.dataset.m;play()})});
$(‘#replay’).addEventListener(‘click on’,play);
operate fillBars(root){$$(‘.bar .t i’,root).forEach(operate(b){b.type.width=”0″;setTimeout(operate(){b.type.width=b.dataset.w+’%’},scale back?0:80)})}
onShow.p2=operate(){play();fillBars($(‘#p2’))};
onShow.p3=operate(){fillBars($(‘#p3’))};
finalState();$$(‘#p2 .bar .t i, #p3 .bar .t i’).forEach(operate(b){b.type.width=b.dataset.w+’%’});
/* 4 value */
var PI=0.15,PO=0.47,PC=0.016;
operate cash(x){return ‘$’+x.toLocaleString(‘en-US’,{minimumFractionDigits:2,maximumFractionDigits:2})}
operate value(){
var I=+$(‘#cin’).worth,O=+$(‘#cout’).worth,C=+$(‘#cch’).worth/100;
var a=I*(1-C)*PI,b=I*C*PC,c=O*PO,t=a+b+c;
$(‘#vin’).textContent=I+’M’;$(‘#vout’).textContent=O+’M’;$(‘#vch’).textContent=Math.spherical(C*100)+’%’;
$(‘#ctot’).textContent=cash(t);
$(‘#s1′).type.width=(a/t*100)+’%’;$(‘#s2′).type.width=(b/t*100)+’%’;$(‘#s3′).type.width=(c/t*100)+’%’;
$(‘#l1’).textContent=”Fresh input “+cash(a);$(‘#l2’).textContent=”Cached input “+cash(b);$(‘#l3’).textContent=”Output “+cash(c);
}
[‘cin’,’cout’,’cch’].forEach(operate(id){$(‘#’+id).addEventListener(‘enter’,value)});
value();
/* resize */
window.addEventListener(‘load’,postH);window.addEventListener(‘resize’,postH);
if(window.ResizeObserver){new ResizeObserver(postH).observe(doc.getElementById(‘qwo38’))}
setTimeout(postH,300);
})();

