通义千问发布 Qwen3.8-Omni-Flash：1M 上下文的全模态模型，主打智能体式音视频理解和工具调用Qwen Releases Qwen3.8-Omni-Flash: a 1M-Context Omni Model Built for Agentic Audio-Video Understanding and Tool Calling

Alibaba's Qwen team released Qwen3.8-Omni-Flash, which it calls its first omni model built around agentic capability. It takes in text, images, audio and video, and still emits text. Audio-video understanding, reasoning and tool calling all sit inside a single model. The team describes its workflow simply: understand the content first, then plan the task, then call tools to execute, and finally deliver the result.

It takes in text, images, audio and video, and emits text. Audio-video understanding, reasoning and tool calling all live in one model. The official workflow is four steps: understand the content, plan the task, execute with tools, deliver the result. At launch it's API-only, with no weights released.

The One-Line Summary

Alibaba's Qwen team released Qwen3.8-Omni-Flash, which it calls its first omni model built around agentic capability. It takes in text, images, audio and video, and still emits text. Audio-video understanding, reasoning and tool calling all sit inside a single model. The team describes its workflow simply: understand the content first, then plan the task, then call tools to execute, and finally deliver the result.

Will it ship? Today it's available as a hosted API. It's already on Qwen Cloud, Alibaba Cloud Model Studio and Qwen Studio. No weights were released at launch, so self-hosting isn't an option for now.

What Qwen3.8-Omni-Flash Is

It builds on the Qwen3.8-Flash-Next architecture, a base model released with open weights in August 2026.

The context window is 1M tokens. Qwen Cloud lists a maximum input of 991K, a maximum output of 131K, and a maximum reasoning length of 262K tokens.

Output is text only. The Model Studio docs remind developers: use Qwen3.5-Omni if you need speech generation. Thinking is on by default with reasoning_effort set to xhigh; setting it to none turns thinking off.

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

On Long Video, It Changed Its Approach

Most video models take a long file and scan it front to back — even when the answer only lives in three minutes of footage somewhere inside.

The Qwen research team describes a different route: start from the question, let the agent decide which segments to watch and which to listen to, then gather evidence over several passes from coarse to fine. Compute and tokens go to the segments that actually matter.

The team's published OmniVideoBench result: accuracy rises from 63.4 to 67.8, while token usage drops from 145,736 to 79,117 — about 45.7% less.

The Scorecard From the Team

All the numbers below come from Qwen itself. As of publication, no independent reproduction has been seen.

• Across 29 benchmarks, the average score is more than 25% higher than Qwen3.5-Omni-Plus
• WildClawBench-MM is up 36.5 points, and AgenticVBench is up 22.3 points
• UniClawBench scores 69.6
• LongAudioSpan is up 8.3 points, and OmniVideoBench is up 9.6 points
• On OmniCap-IF, CSR and ISR are up 8.5 points and 14.1 points respectively

The research team says audio-video performance is now close to Gemini 3.8 Flash, and overall audio performance surpasses it. The official tweet sums up the agentic gains in one line: an average of +19.5 points on WildClawBench-MM and UniClawBench.

Pricing and a Few Hard Limits

Qwen Cloud pricing is $0.15 per million input tokens and $0.47 per million output tokens, with implicit cache hits billed at $0.016 per million tokens.

Compared with Qwen3.5-Omni-Plus, the team reports steep cost reductions: audio input is more than 98% cheaper per hour, and audio-video input is more than 93% cheaper per hour. The tweet puts video input at roughly 89% lower.

A few limits from the Model Studio docs worth remembering:

• Video is passed as a link, up to 2 hours long and up to 2GB
• Audio can run up to 3 hours
• Audio input covers 113 languages and dialects
• Results are stable when video is sampled at 15 fps
• use_multichannel supports two-channel stereo and four-channel FOA spatial audio
• Available in 6 regions: Beijing, Singapore, Hong Kong (China), Tokyo, Frankfurt and Virginia

Calling it with the OpenAI SDK takes just a few lines:

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="")

Companion Open Source: Qwen-MM-Plugins and Qwen-Live Harness

The model only emits text, so handling media has to be handed off to tools. The Qwen team open-sourced two projects for that.

Qwen-MM-Plugins is already live under Apache-2.0, with the tagline "make any agent harness multimodal by default." Each capability is packaged as a Skill plus an optional MCP server. Its guided installer supports Claude Code, CodeBuddy, Codex, Qoder, OpenClaw, Qwen Code and Gemini CLI.

The Omni line of capabilities maps to several demos from the launch:

• omni-memory: build an audio-video memory for a long video
• omni-video2note: turn a tutorial video into an illustrated PDF
• omni-chatcut: music-to-MV, film and TV commentary, and video translation that preserves the speaker's voice

There's also a core plugin that lets the main model read local images and video frames directly. The README also admits a current gap: most harnesses still can't feed audio natively to the main model, so audio has to go around the API for now.

Key Takeaways

• Qwen3.8-Omni-Flash takes text, images, audio and video, and emits text
• 1M-token context, with function calling and web search support, and thinking on by default
• With agentic perception, OmniVideoBench rises from 63.4 to 67.8 while using about 45.7% fewer tokens
• Qwen Cloud pricing: $0.15 per million input tokens, $0.47 per million output tokens
• At launch it's API-only, plus Apache-2.0 Qwen-MM-Plugins to hook into various harnesses

This is a complete translation of the original article, rewritten sentence by sentence into natural English, with images taken from the original and supplemented by public screenshots from media coverage and project repositories. Original author: Asif Razzaq, published September 18, 2026. Benchmark scores and pricing are translated as stated in the original, vendor-reported figures are flagged in the body text, and nothing has been added or removed.