Alibaba Updates Multimodal Generative Model Family at Yunqi Conference

Looking at this Qiwu Conference's announcements together, what Alibaba wants to tell is not three isolated product updates, but an evolutionary main thread: the center of value for generative models is shifting from "one-off content output" toward "a creative process that can be further processed, organized, and controlled." Images are described as moving toward editable visual assets, music toward finer control over melody, aesthetics, and creative intent, and video toward longer narratives, more complex conditions, and further exploration of real-time streaming generation.

This main thread also explains why Alibaba both released specific models and offered a farther-reaching judgment at the same conference. Zheng Bo said on site at the Qiwu Conference that within three years a native omni-modal unified model will appear, and experience will no longer be limited by modality boundaries. It must be made clear that this is a predictive judgment from Alibaba's side, not an already-achieved technological fact.

Supporting this main thread is the capability accumulation of the past year or more. Wan 2.6, released at the end of 2025, already supports intelligent storyboard scheduling, reference-to-video, multi-character consistency, and other capabilities; video generation began moving from text-to-video and image-to-video toward creation jointly controlled by more reference information. Happy Horse 1.0, released this April, further adopted a single-stream Transformer to natively generate audio and video jointly. These are Alibaba's statements about its own model evolution path.

The image portion of the release is Qwen-Image-3.1. Alibaba's problem framing is direct: once entering production scenarios, an image looking good is not enough; creators need to complete tasks stably and controllably, rather than repeatedly "drawing cards" and relying on probability to stumble onto a usable image. The corresponding requirements are accurate information, refined layout, understanding of professional content, and results that ideally can be used directly.

Alibaba illustrates its capabilities with three types of scenarios: scientific illustrations, marketing images, and film storyboards. The generated image shown on site with the theme "A Coffee's Flavor Journey" is said to simultaneously handle Chinese titles, small text, diagrams, and paragraphs, and organize a clear visual hierarchy. In the film storyboard scenario, given a plot segment, the model can generate a storyboard script containing visuals, dialogue, action, and camera language.

A more specific capability description is the editing dimension. Qwen-Image-3.1 provides editing capabilities across five dimensions—"point, line, plane, volume, field"—corresponding respectively to pixel anchoring, structure outlining, layout construction, character shaping, and world building. Zheng Bo's statement is that editing capability reflects, to a certain extent, the reasoning capability of an image model—editing here is not only local modification, but also includes understanding and extracting structure from an existing image, then reorganizing and generating along different dimensions.

The scientific research scenario was singled out because it has a clear precedent for failure. In 2024, Frontiers in Cell and Developmental Biology retracted a paper containing AI-generated figures, in which severely distorted rat anatomical structures and incorrect labels appeared. From this, Alibaba proposes that when facing academic schematic diagrams, models need to achieve "rendering of world knowledge": first understand the scientific principle, then accurately present concepts, processes, and relationships. This also means that the prerequisite for AI-generated content entering scientific research production is that authors verify accuracy and clearly disclose AI use.