--- title: "Alibaba's new model listens, watches, and starts working" date: 2026-10-22 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-qwen38-omni-flash language: en --- # Alibaba's new model listens, watches, and starts working > Alibaba's Qwen3.8-Omni-Flash takes text, image, audio and video together and is built to plan tasks and call tools, not just summarize. The demo starts with a problem most office workers know by heart: a long video meeting, a mountain of footage, and no time to watch it. A Chinese lab now says its newest release can take that recording and turn it into a written summary, a set of action items, and a short edited clip — without a human touching the timeline. For years the flashy models treated sound and moving images as something to describe after the fact. The new one is built to act on them. Alibaba's Qwen team calls the system Qwen3.8-Omni-Flash, and it sits at the center of a quiet but important shift in how Chinese developers think about the large model (大模型). ## One model, four kinds of input Most consumer chat tools still separate the senses. You paste text, or you upload a picture, or you speak into a microphone. Qwen3.8-Omni-Flash is what researchers call a native omnimodal system: text, image, audio, and video enter the same model at the same time, and it keeps a context window of one million tokens — enough to hold about an hour of continuous audiovisual input in a single session. Hold a million-token window and the model can keep an entire lecture, a full codebase, or a long documentary in memory while it works — context that used to require stitching files together by hand. That matters because real work is rarely one sense at a time. A product manager screenshares a buggy app; a teacher records a lesson; a factory supervisor films a stalled machine. The model is designed to take those streams and respond with something finished, not a caption. ## From "understand" to "do" The Qwen team is explicit about the goal. In its own words, the point is to move omnimodal systems from "understanding content" to "planning tasks, calling tools, and completing creative work." In plain terms: the system is meant to behave like an agent (智能体) that can run a workflow, not a chatbot that answers questions. Concretely, Alibaba points to a handful of workflows: - video editing and music-video creation from a single prompt - film production support and spoken commentary - audiovisual summarization for long meetings - real-time conversation that reacts to what it sees and hears The agent (智能体) framing is the throughline. The model is not merely describing a video; it is supposed to open the right tool, run the step, and hand back a file the user can use. ## The numbers Alibaba is publishing Alibaba reports a sizable jump over its previous omnimodal release. Across about thirty evaluations, the company says the new model's average score improved by more than 25% compared with Qwen3.5-Omni-Plus. On agent (智能体)-style audiovisual tests, it claims gains of 36.5 points on WildClawBench-MM and 22.3 points on AgenticVBench, with a score of 69.6 on UniClawBench. Price is part of the story. Alibaba states that the cost of audio input through its interface fell by more than 98% per hour, and the cost of combined audiovisual input fell by more than 93%. On speaker recognition, it reports cutting error rates on the AliMeeting test from 88.11 to 3.35 on one measure and from 89.61 to 17.18 on another. These are Alibaba's own benchmarks and its own pricing, not independent audits. They tell you the direction the team is pushing, not a neutral scoreboard. ## Where it runs The model is delivered through the cloud rather than as downloadable weights, which means developers meet it as a hosted service on Alibaba's Qianwen platform and on Alibaba Cloud. The company says the service is available across regions including Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia. That distribution is a small but real signal. A Chinese-built omnimodal model is now reachable by developers in Europe, North America, and East Asia through the same endpoint — a reminder that the race to power the next wave of agents (智能体) is global, not local. ## Why it matters outside China Two things travel well beyond the demo reel. First, the agent (智能体) emphasis: Chinese labs are betting that the next useful product is a model that finishes a task, not one that writes about it. Second, the price collapse on audio and video input lowers the cost of building tools that actually watch and listen — useful for accessibility, meeting notes, and field work. The shift also changes who can compete. When audio and video stop being premium-priced inputs, smaller teams can build meeting tools, field-inspection helpers, and classroom aids without a research budget. That is the part worth watching. For global readers, the practical question is not "is this better than model X" but "what can I build that used to be too expensive to attempt." ## Honest limitations The benchmark gains and price cuts above come from Alibaba's own reports, and no independent lab has reproduced them yet. The model is cloud-hosted, so teams that need full weight access for private deployment cannot get it here. Real-time, long-horizon audiovisual agents still depend on tooling and runtime support that is early-stage. Independent testing of accuracy on languages other than English, and on Chinese dialects, remains limited. ## What readers can do now - Try the hosted endpoint on Alibaba Cloud with a short internal video and compare the summary against a human-written one before trusting it. - If you build agents (智能体), prototype one audiovisual workflow — a meeting recap or a lecture index — and measure the cost per run. - Watch whether independent benchmarks publish scores for Qwen3.8-Omni-Flash before adopting it for production. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-qwen38-omni-flash Free to quote with attribution and a link to the original.