Alibaba has just launched Qwen 3.8 Omni Flash, its next-generation omnimodal model supporting text, image, audio, and video inputs with a 1-million-token context window. This is not merely a new multimodal model, but a clear shift in direction: from describing audio-video content to planning, tool calling, and executing creative tasks.
Performance and Pricing Highlights
Across 29 benchmarks, Qwen 3.8 Omni Flash improved average scores by more than 25% compared to Qwen 3.5 Omni Plus. Notably, audio input API pricing dropped by over 98%, and audio-video pricing decreased by more than 93% per hour. This cost reduction completely transforms the economics for always-on applications like real-time voice assistants.
In multi-speaker speech recognition, the model delivered impressive results on AliMeeting: a 3.4% DER and 17.2% cpWER, compared to 72.6% and 53.1% for Gemini 3.8 Flash.
Agentic Understanding: Searching Videos Instead of Watching Them in Full
The most notable technical feature is the Agentic Long-Form Audio-Visual Understanding mode. Instead of processing an entire multi-hour video from start to finish, the model starts from the user’s query, autonomously decides which parts to watch and listen to, and gathers evidence across multiple refinement rounds.
On OmniVideoBench, this approach increased accuracy from 63.4% to 67.8% while reducing token consumption from 145,736 to 79,117—a drop of about 45.7%. If independently verified, these results demonstrate that changing how a model searches for information can be more effective than simply ramping up processing.
From Understanding to Action
Alibaba describes Qwen 3.8 Omni Flash as a transition from “multimodal content perception” to “task planning, tool calling, and creative execution.” Concrete demos illustrate this: Music2MV analyzes rhythm, mood, and lyrics to plan an entire music video; a translation pipeline handles multi-speaker speech recognition, translation, voice cloning, and dubbing within a single workflow.
Notably, Alibaba used Qwen 3.8 Omni Flash itself to improve a smaller model for Sichuanese speech recognition. The larger model autonomously selected benchmarks, established baselines, listened to error audio samples, generated 3,413 training examples across four trial rounds, and reduced the character error rate from 25.79% to 15.30% in under 12 hours—with no human intervention.
Accompanying Tools
Two open-source tools debuted alongside the model. Qwen-MM-Plugins is a suite of plugins that integrates into existing agent harnesses (Codex, Claude Code, Gemini CLI, OpenClaw) to add multimodal capabilities. Qwen-Live Harness is a real-time runtime for continuous interaction, supporting task delegation, proactive interaction, and long-term memory management.
It is worth noting that all benchmark comparisons with Gemini 3.8 Flash and DeepSeek-V4-Flash were self-evaluated by Alibaba. No independent benchmarks had been published at the time of launch. Empty cells in the comparison table also indicate that the evaluation matrix remains incomplete.