Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use

Alibaba’s Qwen team has released Qwen3.8-Omni-Flash. They called it its first omni-modal model built around agentic capabilities. It accepts text, images, audio, and video, and it returns text. Audio-video understanding, reasoning, and tool use sit inside one model. The stated workflow is simple: understand the content, plan the task, execute with tools, deliver the result.

Is it deployable? Yes, as a hosted API today. It is live on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio. No open weights were announced at launch, so self-hosting is not an option.

SK Hynix Debuts Ventures CVC Brand at Inaugural Silicon Valley Event – Unite.AI

NASA’s Moon Orbiter Has Spotted An Impact Crater That Only Happens Once A Century

What is Qwen3.8-Omni-Flash

The Qwen3.8-Omni-Flash is built on the Qwen3.8-Flash-Next architecture. That base model shipped with open weights in August 2026.

The context window is 1M tokens. QwenCloud lists 991K max input and 131K max output. Max reasoning length is 262K tokens.

Output is text only. The Model Studio docs point developers to Qwen3.5-Omni when they need generated speech. Thinking is on by default, with reasoning_effort set to xhigh. Setting it to none disables thinking.

The API follows both the DashScope and OpenAI protocols. It works with Chat Completions and the Responses API. Function calling, web search, structured outputs, context caching, and batch calls are supported.

Agentic Perception for Long Video

Most video models read a long file from start to finish. That holds even when the answer sits in 3 minutes of footage.

Qwen research team describes a different path. The agent starts from the question. It decides what to watch and hear. It then gathers evidence over several coarse-to-fine rounds. Compute and tokens go to the segments that matter.

The research team reports the result on OmniVideoBench. Accuracy rises from 63.4 to 67.8. Token use drops from 145,736 to 79,117. That is about 45.7% fewer tokens.

Reported Benchmarks

All figures here come from Qwen. Independent results were not available at publication.

Across 29 evaluations, the average score improves more than 25% over Qwen3.5-Omni-Plus.
WildClawBench-MM improves by 36.5 points. AgenticVBench improves by 22.3 points.
UniClawBench reaches 69.6.
LongAudioSpan gains 8.3 points. OmniVideoBench gains 9.6 points.
OmniCap-IF CSR and ISR improve by 8.5 and 14.1 points.

The research team states that audio-visual performance is close to Gemini 3.8 Flash. It also claims overall audio performance above Gemini 3.8 Flash. The X post summarizes the agent gains as +19.5 points on average across WildClawBench-MM and UniClawBench.

🚀 Meet Qwen3.8-Omni-Flash, Qwen’s first omni-modal model built around agentic capabilities!

Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result.

Highlights: 🥳
-… pic.twitter.com/iJypeohw7y

— Qwen (@Alibaba_Qwen) September 18, 2026

Pricing and Input Limits

QwenCloud lists $0.15 per 1M input tokens and $0.47 per 1M output tokens. Implicit cache hits cost $0.016 per 1M tokens.

The research team reports large cost cuts against Qwen3.5-Omni-Plus. Audio input costs over 98% less per hour. Audio-visual input costs over 93% less per hour. The X post puts the video input reduction at about 89%.

Key limits from the Model Studio docs:

Video files up to 2 hours and 2 GB by URL.
Audio files up to 3 hours.
Audio input in 113 languages and dialects.
Stable results with video sampled at up to 15 fps.
Two-channel stereo and four-channel FOA spatial audio through use_multichannel.
Availability in 6 regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia.

Calling it takes a few lines with the OpenAI SDK:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DASHSCOPE_API_KEY"],
    base_url=os.environ["DASHSCOPE_BASE_URL"],
)
completion = client.chat.completions.create(
    model="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.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

The model returns text, so tools do the media work. Qwen team is open-sourcing 2 projects to support that.

Qwen-MM-Plugins is live under Apache-2.0. Its tagline is ‘Make any agent harness multimodal-native.’ Each capability installs as a Skill plus an optional MCP server. The guided installer supports Claude Code, CodeBuddy, Codex, Qoder, OpenClaw, Qwen Code, and Gemini CLI.

The Omni capabilities map to the launch demos:

omni-memory builds an audio-visual memory of a long video.
omni-video2note converts a tutorial video into an illustrated PDF.
omni-chatcut covers Music-to-MV, movie commentary, and speaker-preserving video translation.

A core plugin lets the main model read local images and video frames natively. The README notes one current gap. Most harnesses cannot feed audio to the main model natively yet. Audio is routed through the API for now.

Interactive Explainer

‘+x+”}).join(”); $$(‘#loop .node’).forEach(function(n,i){n.classList.toggle(‘on’,i===st);n.classList.toggle(‘done’,i<st function="" stop="" the="" loop="" .chip="" .node="" if="" st="0;drawLoop();$('#run').textContent="Pause";" timer="setInterval(function(){if(st">=3){stop();return}st++;drawLoop();postH()},reduce?900:1900); }); $(‘#rst’).addEventListener(‘click’,function(){stop();st=0;drawLoop()}); drawLoop(); /* 2 perception */ var tl=$(‘#tl’),cells=[];for(var i=0;i=plan.length){finalState();return} var p=plan[n];cells[p[0]].className=p[1];$(‘#stg’).textContent=p[2]; n++;var f=n/plan.length;setK(F.tok*f,F.acc*f); setTimeout(tick,step); })(); } $$(‘#mode button’).forEach(function(b){b.addEventListener(‘click’,function(){ $$(‘#mode button’).forEach(function(x){x.classList.toggle(‘on’,x===b)});mode=b.dataset.m;play()})}); $(‘#replay’).addEventListener(‘click’,play); function fillBars(root){$$(‘.bar .t i’,root).forEach(function(b){b.style.width=”0″;setTimeout(function(){b.style.width=b.dataset.w+’%’},reduce?0:80)})} onShow.p2=function(){play();fillBars($(‘#p2’))}; onShow.p3=function(){fillBars($(‘#p3’))}; finalState();$$(‘#p2 .bar .t i, #p3 .bar .t i’).forEach(function(b){b.style.width=b.dataset.w+’%’}); /* 4 cost */ var PI=0.15,PO=0.47,PC=0.016; function money(x){return ‘$’+x.toLocaleString(‘en-US’,{minimumFractionDigits:2,maximumFractionDigits:2})} function cost(){ var I=+$(‘#cin’).value,O=+$(‘#cout’).value,C=+$(‘#cch’).value/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.round(C*100)+’%’; $(‘#ctot’).textContent=money(t); $(‘#s1′).style.width=(a/t*100)+’%’;$(‘#s2′).style.width=(b/t*100)+’%’;$(‘#s3′).style.width=(c/t*100)+’%’; $(‘#l1’).textContent=”Fresh input “+money(a);$(‘#l2’).textContent=”Cached input “+money(b);$(‘#l3’).textContent=”Output “+money(c); } [‘cin’,’cout’,’cch’].forEach(function(id){$(‘#’+id).addEventListener(‘input’,cost)}); cost(); /* resize */ window.addEventListener(‘load’,postH);window.addEventListener(‘resize’,postH); if(window.ResizeObserver){new ResizeObserver(postH).observe(document.getElementById(‘qwo38’))} setTimeout(postH,300); })(); </st>

Credit: Source link