Qwen3.8-Omni-Flash: omnimodal agents at a fifth of Gemini Flash
API-only, it reads audio, video and images at 1M context, bills audio under a cent an hour and runs live camera agents — plus Cua's tiny open form filler.
Copy markdown
A fifth of Gemini Flash for the same omnimodal job
Alibaba's Qwen3.8-Omni-Flash bills $0.15 in / $0.47 out per million tokens (cache hits $0.016) against Gemini 3.8 Flash's $0.75 / $3.75 — roughly 5x cheaper on input and 8x on output. It ingests text, images, audio in 113 languages and video up to two hours, all into a 1M-token context, and returns text.
Audio for under a cent an hour
Audio processing costs over 98% less than Qwen3.5-Omni-Plus — under $0.01 per hour — and 720p video at 1 fps runs about $0.20/hr. That makes always-listening transcription, meeting notetakers and call-analytics agents cheap enough to just leave running.
Build this weekend: a real-time camera-and-mic agent
It ships agentic perception (coarse-to-fine passes that zoom on question-relevant video segments), function calling, web search and structured outputs, and it's OpenAI-SDK compatible. Point a phone camera at it for a live assistant, auto-edit a vlog, or summarize a two-hour recording. It averages 25%+ over Qwen3.5-Omni-Plus across 29 evals (OmniVideoBench 63.4 to 67.8).
The catch: no open weights
It's API-only at launch — QwenCloud, Alibaba Cloud Model Studio and Qwen Studio — with no weights released, so there's no local or offline path. The open-source pieces (Qwen-MM-Plugins and a Qwen-Live harness) cover integration, not the model itself.
Elsewhere: Cua open-sources a 706K-param form filler
Cua released CUA-S1-FORMS, a byte-level classifier that scores bounded form actions — fill, check, click, skip — in a single forward pass instead of calling an LLM, reportedly around 99.95% top-1 on synthetic tests from a 2.8 MB checkpoint. It's MIT-licensed; reports differ on whether the weights shipped or just the training code and dataset. Aimed at cheap, deterministic form automation.