For years the flashy "reasoning model (推理模型)" race was a contest of text against text. A model read a hard math problem and talked its way to an answer. A team at the Chinese Academy of Sciences decided that was not enough — they wanted a model that reasons while it is still looking at the picture, hearing the clip, or reading the sensor feed.
Key takeaways
- Zidong Taichu (紫东太初) is a multimodal large model (大模型) developed by the CAS Institute of Automation (中科院自动化所) with the Wuhan AI Research Institute (武汉人工智能研究院).
- Version 4.0 was unveiled on 19 September 2025 at the Donghu International AI Summit and is described by its developers as the world's first "deep-reasoning + multimodal" fusion model.
- The line has iterated four times since its 1.0 release in July 2021, when it was billed as the world's first 100-billion-parameter multimodal large model (大模型).
- 4.0 introduces a "see, recognize, think" (边看、边识、边思) loop: it can crop, zoom, rotate, and locate regions inside an image as part of its reasoning, not after it.
- The team has open-sourced ZDTaichu5.0-9B, a 9-billion-parameter checkpoint, and the earlier multimodal model won the WAIC SAIL Award in 2022.
From perception to reasoning
The project began as a cross-modal platform rather than a chatbot. Version 1.0, shown at the World Artificial Intelligence Conference in July 2021, combined text, image, and speech in a single model and was presented as the first 100-billion-parameter multimodal large model (大模型). Version 2.0, described in 2023, added 3D, video, and sensor-signal modalities under a "full-modal" banner and was built on a domestic hardware stack. By 4.0, released in September 2025, the emphasis had shifted from perceiving the world to reasoning about it.
The work is led by Wang Jinqiao (王金桥), a researcher at the CAS Institute of Automation and dean of the Wuhan AI Research Institute. The center behind the model was established in 2022 as an independent research unit of the institute, framing the effort as a state-backed but research-driven program rather than a commercial product launch.
What "reasoning while watching" means
The 4.0 release describes a three-stage evolution: pure-text thinking, simple image-grounded thinking, and then fine-grained multimodal interleaved thinking. The leap is that the model operates on visual content during its reasoning chain — translating, zooming, rotating, locating, enhancing, or reconstructing regions of an image — instead of first describing an image and then reasoning about the description.
The developers demonstrate this with concrete tasks. In video understanding, they cite the ability to precisely locate a segment within a 180-minute long video and summarize it. In audio understanding, a spoken request such as "I want to book a respiratory department appointment" is shown triggering an app operation that selects the right clinic based on symptoms. These are the team's own demonstration scenarios, but they illustrate the design goal: a model that plans and acts across modalities the way a person does when handling a real-world task.
The open piece
Alongside the flagship, the team open-sourced ZDTaichu5.0-9B, a 9-billion-parameter checkpoint, giving outside developers a smaller, deployable entry point to the architecture. The earlier multimodal model earned the WAIC SAIL Award in 2022, one of the conference's top research honors, which helped establish the line's credibility before the reasoning-focused 4.0.
The launch also introduced Zidong Taichu Cloud (紫东太初云), described by the organizers as the first multimodal large-model native collaborative cloud and the first full-stack domestic (全栈国产化) 10,000-card intelligent-computing cloud. At the 4.0 event, 28 enterprises signed on as ecosystem partners, and the institute has reported partnerships with more than 1,000 enterprises and around 60 applied use cases across dozens of industries.
Why it matters for China's model map
Most of the public "reasoning model (推理模型)" conversation has centered on text-only systems that excel at math and code. Zidong Taichu argues for a different axis: reasoning that stays fused with vision, audio, and signals. That orientation maps onto Chinese industrial priorities — smart manufacturing, smart healthcare, autonomous driving, and low-altitude economy — where a model must interpret a camera feed or a sensor stream, not just a prompt.
The "full-stack domestic" framing also matters in a market where compute and tooling sovereignty are explicit policy goals. A multimodal reasoning model paired with a domestically sourced training-and-inference cloud is positioned as a vertically integrated option for regulated or sovereignty-sensitive deployments.
Honest limitations
- The strongest performance claims, including comparisons to named frontier models, come from the developers and a Wuhan district government release; I have not found independent head-to-head benchmark reports, so I treat capability claims as the team's own and keep them qualitative.
- The 4.0 launch and partnership figures are drawn from the institute's site and a local government notice; partnership and application counts are the organizers' own and are not independently audited.
- I did not independently verify the exact parameter count or training-data volume of 4.0 beyond the open-sourced 9B checkpoint; those specifics were not published in the sources I could confirm.
- This piece focuses on the multimodal reasoning direction and does not cover the team's embodied or agent-specific work in depth.
What readers can do now
- If you build vision- or audio-grounded agents, test the open ZDTaichu5.0-9B checkpoint on a hard image-grounded reasoning task — such as long-video segment localization — before assuming a text-only model will do.
- Watch the Zidong Taichu Cloud (紫东太初云) full-stack domestic compute offering if you need a China-hosted, domestically sourced multimodal stack for regulated or sovereignty-sensitive workloads.
- Separate developer demos from evidence: when evaluating any multimodal reasoning model, run your own task suite rather than quoting a launch-stage comparison.