Qwen3.8-Omni-Flash: 1M-context audio-video API, audio 98% cheaper
Alibaba's omni model takes text, images, audio and video into 1M tokens and calls tools — at $0.15/$0.47 per M, but API-only with no speech output.
Copy markdown
One model for text, images, audio, and video
Qwen3.8-Omni-Flash ingests all four modalities into a 1M-token window (991K in, 131K out) and returns text, with function calling, web search, and structured outputs built in. Point one endpoint at a video, a PDF, or a call recording instead of chaining three models.
Audio drops 98%, audio-visual 93%
Pricing is $0.15 per million input tokens, $0.47 output, and $0.016 on cache hits. Alibaba says audio runs 98% cheaper and audio-visual 93% cheaper than Qwen3.5-Omni-Plus — cheap enough to leave a transcription or video-watching agent running all day.
It watches video without reading every frame
'Selective perception' picks the relevant segments instead of ingesting the whole file. On OmniVideoBench, accuracy climbed from 63.4 to 67.8 while using about 46% fewer tokens — the gap between a video agent you can afford and one you can't.
+25% over its predecessor, near Gemini on A/V
Qwen claims a 25% average gain across 29 evals versus Qwen3.5-Omni-Plus (+36.5 on WildClawBench-MM, +22.3 on AgenticVBench), with audio-visual quality close to Gemini 3.8 Flash and overall audio above it. Thinking is on by default at xhigh effort.
The catch: API-only, no speech out
No open weights at launch — it's proprietary via QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio across six regions (incl. Frankfurt and Virginia). It only returns text; if you need generated speech, you're still routing to Qwen3.5-Omni.