Alibaba (Qwen) · 2026-09-18 · major
Qwen3.8-Omni-Flash — Alibaba's omni model gets a 1M-token context
Qwen3.8-Omni-Flash is Alibaba's omnimodal model that reads text, images, audio and video in a 1M-token context. Alibaba says audio input costs over 98% less per hour than Qwen3.5-Omni-Plus, and average scores rose 25% across 29 tests.
Alibaba's omnimodal Flash model takes audio and video in a 1M-token context, at a fraction of the old audio price.
Key specs
| Audio input price | over 98% lower |
|---|---|
| Agent benchmark gain | +19.5 points |
Quick facts
| Maker | Alibaba (Qwen) |
|---|---|
| Input | Text, image, audio and video |
| Output | Text |
| Context window | 1M tokens |
| Weights | Cloud-hosted only |
| Availability | Qwen Chat, QwenCloud, Model Studio API |
| Companion repo | Qwen-MM-Plugins (Apache-2.0) |
What is it?
Qwen3.8-Omni-Flash accepts text, images, audio and video in one 1M-token context window and answers in text. Alibaba released it on 18 September 2026 through Qwen Chat, QwenCloud and the Model Studio API, on an OpenAI-compatible endpoint. Thinking mode, tool calling and web search are all supported, so one call can watch a clip and then act on what it saw.
How does it work?
Alibaba describes the model as natively omnimodal: a single model handles all four input types rather than a chain of separate converters feeding a text model. Efficiency is where the work went — Alibaba reports 51.8% fewer tokens on OmniVideoBench at comparable accuracy, which is most of where the price drop comes from. Alongside the model, Alibaba published Qwen-MM-Plugins under Apache-2.0, a package that lets text-only agent harnesses accept audio and video.
Why does it matter?
Cost is the real change. Audio input drops more than 98% per hour against Qwen3.5-Omni-Plus, and audio-visual input more than 93%, which moves meeting summaries, video captioning and long-form transcription out of the too-expensive-to-run bracket. Alibaba says average scores across 29 evaluations rose more than 25% over that predecessor, with a 19.5-point average gain on agent tests. The pricing gap against Gemini was the main topic of a 196-point Hacker News thread.
Who is it for?
teams building audio and video agents
Frequently asked questions
- Are the Qwen3.8-Omni-Flash weights open?
- No. Qwen3.8-Omni-Flash is served from the cloud only, through Qwen Chat, QwenCloud and the Alibaba Cloud Model Studio API, so there is no weight download the way there is for the open Qwen3.8 models. The one open piece shipped with it is Qwen-MM-Plugins, an Apache-2.0 Python package on GitHub that adds audio and video handling to existing agent harnesses.
- Can Qwen3.8-Omni-Flash speak its answers back?
- No. Alibaba Cloud's documentation lists Qwen3.8-Omni-Flash as taking audio and video in and returning text out. Speech generation stays with the older Qwen3.5-Omni, which answers in both speech and text but is limited to the Beijing and Singapore regions. If a project needs spoken replies, Qwen3.8-Omni-Flash has to be paired with a separate text-to-speech step.
- Which regions can I call Qwen3.8-Omni-Flash from?
- Alibaba Cloud lists six regions for Qwen3.8-Omni-Flash: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia. Each region needs its own API key and workspace URL, so an application has to point at the endpoint for the region it registered in. That spread is wider than Qwen3.5-Omni, which Alibaba Cloud offers in Beijing and Singapore only.
- How does Qwen3.8-Omni-Flash compare with Gemini 3.8 Flash?
- Alibaba says Qwen3.8-Omni-Flash approaches Gemini 3.8 Flash on audio-video benchmarks rather than beating it, and sells the difference on price instead. That framing drove the Hacker News discussion, where commenters weighed the much lower token cost against Gemini's scores. Alibaba has not published a full head-to-head score table alongside the announcement.
Try it
Call model id qwen3.8-omni-flash on the OpenAI-compatible Model Studio endpoint