Qwen3.8-Omni-Flash is Alibaba’s September 18, 2026 omni-modal release. Where Qwen3.8-Flash-Next is an open-weight efficiency preview, Omni-Flash is a production API model with one aim: move omni-modal models from understanding audio and video to planning tasks, calling tools and delivering finished work.
It accepts text, images, audio and video, supports a 1M-token context, and keeps text ability on par with a text-only model of its size. Alibaba reports an average gain above 25% over Qwen3.5-Omni-Plus across 29 evaluations, while API pricing for audio input fell by more than 98% and audio-video input by more than 93% (Alibaba’s own per-hour estimate).
What it is built for
- Long audio-video understanding. On-demand, controllable captions, plus an agentic mode that decides what to watch and listen to. On OmniVideoBench, Alibaba says that mode lifts accuracy from 63.4 to 67.8 while cutting tokens per query from 145,736 to 79,117.
- Meetings. Up to an hour of audio-video with joint speaker diarisation, transcription and identity matching; Alibaba shows it producing minutes, action items and even starting to code from the meeting.
- Production workflows. Music-to-music-video planning, drama translation with character-aware dubbing, long-film commentary and video-to-notes, built with the open-source Qwen-MM-Plugins.
- Realtime. A separate Qwen3.8-Omni-Flash-Realtime API handles live voice and video over WebSocket and WebRTC, with an open-source Qwen-Live Harness; Alibaba calls it the first omni model to support locating a sound source.
Benchmarks
Vendor-reported by Alibaba (September 18, 2026). Comparison columns are as published.
| Benchmark | Omni-Flash | Qwen3.5-Omni-Plus | Gemini 3.8 Flash | Seed 2.0 Lite | Muse Spark 1.2 |
|---|---|---|---|---|---|
| WildClawBench-MM | 71.0 | 34.5 | 58.9 | 41.9 | — |
| UniClawBench | 69.6 | 67.1 | 69.0 | 61.2 | — |
| AgenticVBench | 36.8 | 14.5 | 45.0 | 10.0 | — |
| OmniGAIA | 74.0 | 57.2 | 78.6 | 64.4 | — |
| DailyOmni | 85.1 | 85.1 | 84.0 | 81.4 | 79.6 |
| OmniVideoBench | 63.4 | 53.8 | 65.2 | 58.5 | 62.2 |
| Video-MME-v2 | 65.0 | 47.9 | 71.0 | 64.9 | — |
| StreamingBench | 80.8 | 57.1 | 79.9 | 77.2 | 77.8 |
| SpotSoundBench | 67.2 | 64.2 | 39.7 | 59.6 | 16.9 |
| MuchoMusic-RUL | 72.6 | 71.6 | 53.7 | 61.7 | 40.1 |
| MMAU | 81.8 | 81.9 | 76.9 | 77.2 | 63.5 |
| VoiceBench | 91.6 | 92.9 | 92.3 | 84.1 | 79.8 |
| AliMeeting cpWER (lower is better) | 17.2 | 89.6 | 53.1 | 76.1 | 92.7 |
| FLEURS-ASR WER (lower is better) | 9.3 | 7.2 | 7.9 | 32.1 | 23.6 |
The pattern is consistent with Alibaba’s own framing: audio-video capabilities are “close to Gemini 3.8 Flash” and audio is better overall — while agentic video and long-video reasoning still trail Gemini.
Pricing
| Per 1M tokens | International | Global / mainland China |
|---|---|---|
| Input | $0.15 | $0.113 |
| Cached input | $0.016 | $0.014 |
| Output | $0.47 | $0.382 |
Audio and video are billed as tokens; see Alibaba’s billing rules for how many tokens an hour of media becomes.
Specs
| Spec | Value |
|---|---|
| API models | qwen3.8-omni-flash, qwen3.8-omni-flash-realtime |
| Input | Text, image, audio, video |
| Output | Text (speech via the Realtime API) |
| Context window | 1M tokens |
| Reasoning | Thinking supported; effort xhigh / medium / low |
| Protocols | OpenAI Chat Completions and Responses; DashScope WebSocket / WebRTC for realtime |
| Weights | Not released |
How it compares
- vs Gemini 3.8 Flash: Gemini keeps the lead on agentic video and long-video reasoning; Qwen is stronger on audio grounding, music understanding and multi-speaker meetings.
- vs Qwen3.8-Flash-Next: the open-weight text-and-vision model; Omni-Flash adds audio and is API-only.
- vs Qwen3.8-Max: Max is the 2.4T flagship for reasoning and coding; Omni-Flash is the cost-efficient model for audio-video agents.
Limitations
- API only, with no weights.
- Vendor-reported, and compared mostly against fast-tier models rather than frontier flagships.
- Realtime speech needs the separate Realtime model.
- Multilingual ASR regressed slightly versus the previous generation.
For the wider picture, see the 2026 LLM guide.
Best for
- Video editing, dubbing and film-commentary agents
- Meeting recordings to minutes, speakers and action items
- Music-video planning and timestamped lyric extraction
- Audio-video question answering over long footage
Pros & cons
Strengths
- Takes text, images, audio and video in one model with a 1M-token context — hours of meeting or film footage in a single request
- Alibaba reports an average gain of more than 25% over Qwen3.5-Omni-Plus across 29 evaluations, and API audio-input prices down more than 98%
- Built around agentic audio-video work — WildClawBench-MM 71.0 (up from 34.5) and UniClawBench 69.6 — with open-source Qwen-MM-Plugins and a Qwen-Live Harness
- Strong multi-speaker meeting transcription (AliMeeting DER 3.4 / cpWER 17.2 against 88.1 / 89.6 for its predecessor)
Limitations
- Closed weights and API-only, unlike the sibling Qwen3.8-Flash-Next
- Output is text; speech output comes through the separate Realtime API, not the standard model
- Behind Gemini 3.8 Flash on several audio-visual reasoning rows — AgenticVBench 36.8 vs 45.0, Video-MME-v2 65.0 vs 71.0, OmniGAIA 74.0 vs 78.6
- Multilingual ASR slipped against its predecessor — FLEURS WER 9.3 vs 7.2 (lower is better)
Frequently asked questions
When was Qwen3.8-Omni-Flash released?
Alibaba released it on September 18, 2026 on its Qwen Cloud platform as `qwen3.8-omni-flash`, together with the `qwen3.8-omni-flash-realtime` model for streaming audio and video.
How much does Qwen3.8-Omni-Flash cost?
On the international (Singapore) deployment it is $0.15 per 1M input tokens, $0.016 for cache-hit input and $0.47 per 1M output tokens. The Global and mainland China deployments list $0.113, $0.014 and $0.382.
Are the weights open?
No. Unlike Qwen3.8-Flash-Next, Alibaba has not published weights for Omni-Flash; it is available through the API only.
What can it do that a text-and-image model cannot?
It reads audio and video natively — transcribing multiple speakers, describing a video on demand, locating key moments in hours of footage, and feeding that into tools for editing, dubbing or summarising. Alibaba says its agentic mode cuts tokens per query by about 46% on OmniVideoBench while improving accuracy from 63.4 to 67.8.
Is there a real-time version?
Yes. Qwen3.8-Omni-Flash-Realtime accepts live audio and video streams over WebSocket and WebRTC and can respond and call tools while the user is still speaking. Alibaba also open-sourced the Qwen-Live Harness around it.