Qwen3.8-Omni-Flash is an agent runtime for audio and video, not a talking head
By AgentRiot Editorial
Launched September 14, 2026. The API ID is `qwen3.8-omni-flash`. Speech output is still Qwen3.5-Omni.
Qwen's post puts the new model on the Qianwen platform with text, image, audio, and video in, and a 1M-token context window. The job it wants is plan, call tools, and finish: editing, music videos, film commentary, meeting minutes that turn into tasks.
Alibaba Cloud's Qwen-Omni docs are stricter about output. Use qwen3.8-omni-flash for text analysis of audio and video. Use Qwen3.5-Omni when you need speech generated. Thinking is on by default for the 3.8 Flash ID, with reasoning_effort defaulting to xhigh. Tool calling and web search are in that doc.
What ships with it
Two companion projects sit next to the model. Qwen-MM-Plugins adds on-demand perception and workflow tools for long audio and video, including an open-sourced Video2Note path. Qwen-Live Harness is the open-sourced runtime for continuous, real-time omnimodal interaction. The blog treats model, plugins, and harness as one stack.
API caps, not the marketing graphic: up to 64 files on 3.8 Flash, 2 GB media, 2 hours of audio in one described cap and 3 hours of video in another. Audio tokens are duration_seconds * 7. Read the live table before you size a job. Those two duration lines sit in different parts of the doc.
Vendor numbers
The launch post reports more than 25% average gain versus Qwen3.5-Omni-Plus across 29 listed evals, plus large cuts in estimated per-hour audio and audio-visual input price versus the prior Omni line. Those price cuts use Qwen's own hourly conversion: 30 times the cost of two minutes of source, audio-visual at 720p / 1 fps. They are not a public USD tariff on Alibaba's Qwen-Omni help page.
On agent benches the post cites WildClawBench-MM +36.5, AgenticVBench +22.3, UniClawBench 69.6. It also claims audio-visual performance close to Gemini 3.8 Flash and overall audio above that model. Those rows are vendor-reported.
A separate vendor experiment: the model was tasked with improving Qwen2.5-Omni-3B Sichuan-dialect ASR in 12 hours. Character error rate on WenetSpeech-Chuan fell from 25.79% to 15.30% in that write-up. That is Qwen's account of its own loop, not a third-party replication.
When it is the right API
Use qwen3.8-omni-flash when the job is long audio or video in, text and tool calls out: captions you specify, meeting minutes, Video2Note, an agent that watches only the minutes that matter. The post says agentic understanding on OmniVideoBench went from 63.4 to 67.8 accuracy while cutting tokens per query from 145,736 to 79,117.
Do not use it as a drop-in voice bot. Speech generation is still the 3.5 Omni line.
If you already run Qwen for text agents, this cut is the audio-video worker plus an open harness. Read the live Qwen-Omni doc for the ID, the caps, and the thinking defaults before you wire it.