Qwen model release · Qwen3.8-Omni-Flash explained

Qwen3.8-Omni-Flash turns long audio and video into text and tool calls

The hosted model reads text, images, audio, and video in one request, keeps up to one million tokens of context, and can call functions or search the web. Its documented output is text.

Published by Qwen · Launched September 18, 2026 · Reviewed by Learnetto

For teams processing meetings, calls, and long videos, the practical change is simpler: one model can inspect sound and pictures together, find the relevant moments, and pass a structured action to software. Test it against your current pipeline before replacing transcription, retrieval, or review steps.

Inputs: Text, images, audio, video
Output: Text only
Context: 1 million tokens
APIs: Chat Completions and Responses

What changed

Alibaba Cloud documents Qwen3.8-Omni-Flash as a hosted model for audio and video understanding, meeting summaries, and content analysis. It accepts all four input types through Chat Completions or the Responses API, supports function calling and web search, and uses automatic context caching.

Qwen's launch benchmarks report gains over Qwen3.5-Omni-Plus and competitive results against Gemini 3.8 Flash. Those numbers are vendor-published, and several agent scores measure the model inside a harness. Treat them as candidates for testing rather than proof for your workload.

How the agent pattern works

A long-media agent does not need to feed every frame into one expensive pass. It can inspect a coarse index, choose relevant time ranges, examine those segments more closely, and then return evidence or call a bounded tool. Qwen's companion plugin repository packages versions of that pattern for video memory, editing, search, Blender, FreeCAD, and tutorial generation.

The model remains one part of the system. Your application should own permissions, tool schemas, retries, citations, and final side effects. A plausible transcript or tool argument can still be wrong.

What to build first

Meeting review that uses both speakers' voices and the video feed to draft action items, with names and commitments checked by a person.
Long-video search that returns timestamps and quoted evidence before it writes a summary.
Support-call quality review that separates transcription errors from policy or coaching judgments.
Media triage that classifies an item, then calls a narrow function only when the confidence and evidence meet your threshold.
A multimodal coding-agent tool that inspects a screen recording or design walkthrough while keeping code changes behind normal tests and review.

Test ten recordings before switching

Choose ten recordings your team has already processed and keep the accepted outputs. Run the same files and prompts through the current stack and Qwen3.8-Omni-Flash. Record input and output tokens, latency, speaker mistakes, missed timestamps, unsupported claims, and human correction time.

Move one workflow only if the new route improves the measures that matter. Keep the previous route available while you learn which languages, recording conditions, and task types cause failures.

Limits and open questions

The current international documentation describes text output only. Its model list still points realtime speech conversations to qwen3.5-omni-plus-realtime and does not list a Qwen3.8 realtime speech model. Products needing live voice output still require a separately verified realtime or speech layer.

No open weights were announced for Qwen3.8-Omni-Flash at launch. Qwen-MM-Plugins is Apache-2.0 software, but that does not make the hosted model open-weight. Availability, regional endpoints, prices, and rate limits can also change, so confirm the current Model Studio documentation before committing an architecture.

Audio-video understanding can make confident inferences from sampled frames or incomplete sound. Ask for timestamps and evidence, keep uncertainty visible, and independently check consequential outputs.