Qwen3.8-Omni-Flash

Qwen3.8-Omni-Flash is the Qwen team's first natively omni-modal model built around agentic capabilities, announced on 2026-09-18. It accepts text, images, audio and video in one model with a 1M-token context window, and it reasons over what it sees and hears before calling tools to act on it. Alibaba positions the model for work where a single pipeline step used to mean several different APIs: video editing, music video creation, film narration, multimedia summarization and audio-video dialogue.

Model Specifications

Specification	Qwen3.8-Omni-Flash
Context window	1,000,000 tokens
Max input	991K tokens (983K in thinking mode)
Max output	131K tokens
Max reasoning budget	262K tokens
Input	Text, image, audio, video
Output	Text
Architecture	Built on the Qwen3.8-Flash-Next architecture
Throughput	2M TPM, 30K RPM
Protocols	DashScope and OpenAI-compatible
License	Closed weights

Spatial audio matters here: the model supports two-channel and four-channel spatial audio understanding, which is what makes multi-speaker video and dialogue analysis usable rather than a wall of overlapping text. Alibaba also open-sourced companion Qwen-MM-Plugins so agent frameworks can reach its native multimodal abilities, and it lists a Qwen-Live Harness as coming soon.

Pricing

Rates from the Qwen Cloud model page, checked on 2026-09-23.

Item	Price
Input	$0.15 / 1M tokens
Output	$0.47 / 1M tokens
Implicit cache read	$0.016 / 1M tokens

The Qwen team says video input costs fall by about 89% against Qwen3.5-Omni-Plus, which is the number that matters if you were previously paying separately for a video understanding call and a text model call.

What the release benchmarks claim

These figures come from the official announcement, not from an independent run:

Audio-video agent performance: +19.5 points on average across WildClawBench-MM and UniClawBench.
Agentic perception: 51.8% fewer tokens than static understanding on OmniVideoBench, because the model actively explores long video and locates key moments instead of consuming every frame.
Overall audio-video capability: described as approaching Gemini 3.8 Flash.

Limitations

Closed weights. Qwen3.8-Omni-Flash is an API model; there is no downloadable checkpoint.
Vendor-reported scores. The benchmark deltas above have not been reproduced independently, and "approaching Gemini 3.8 Flash" is a positioning claim, not a measured ranking.
Effective input is 991K, not 1M. The 1M figure is the context window; the practical input cap sits just under it.
Tool use may need the plugins. Alibaba recommends the companion Qwen-MM-Plugins for agent frameworks, so a bare API integration will not expose every multimodal ability.

FAQ

Is Qwen3.8-Omni-Flash the same as Qwen-Audio-3.1?

No. Qwen-Audio-3.1 is a family of audio specialists for transcription, synthesis and realtime speech. Qwen3.8-Omni-Flash is one general model that handles audio and video alongside text and images, and it calls tools instead of only transcribing.

Does it output audio or images?

No. Input is text, image, audio and video, but output is text. For image generation compare Qwen-Image-2.1.

How does it compare to Gemini's omni model?

Alibaba's own claim is that it approaches Gemini 3.8 Flash on audio-video tasks. See Google Gemini Omni Flash for the competing entry.

Alternatives

Qwen3.8-Flash-Next: the text-first architecture this model is built on, if you do not need audio or video.
Google Gemini Omni Flash: Google's omni-modal entry in the same tier.
Qwen-Image-2.1: when the gap is image generation and editing rather than understanding.
Qwen-Audio-3.1: audio specialists when speech is the whole job.