Most labs chased the same recipe after ChatGPT: bolt a language model onto a vision encoder, or spin up a diffusion model for images. One Beijing institute went the other way — and ended up in Nature.
The Beijing Academy of Artificial Intelligence (BAAI, 北京智源人工智能研究院) was founded in 2018, with Huang Tiejun (黄铁军) as its founding director; he now serves as the institute's council chairman. Long before the generative-AI boom, BAAI made an unusual bet: in 2020, two years before ChatGPT, it committed its researchers to the large model (大模型) direction and later, in 2021, popularized the very Chinese term "大模型" for the field.
Key takeaways
- BAAI's Emu3 paper, "Multimodal learning with next-token prediction for large multimodal models," was published in Nature, volume 650, issue 8101, pages 327–333 (DOI 10.1038/s41586-025-10041-x), with Huang Tiejun as a corresponding author.
- The result was reported by Chinese state science media as the first Nature main-journal paper led by a Chinese institution on a large-model topic; it appeared online on 28 January 2026 and in print on 12 February 2026.
- Emu3 uses a single next-token-prediction framework to unify text, image and video understanding and generation, and extends to vision-language-action modelling for robotic manipulation.
- BAAI has released more than 200 open-source models and, in June 2025, launched the "悟界 (Wu Jie)" series aimed at world models, embodied intelligence and AI for Science.
The bet that looked wrong
When BAAI started building large models in 2020, the consensus outside China favored separate, specialized routes: contrastive methods for understanding, diffusion models for generation. A unified autoregressive approach — predicting the next token across every modality — was seen as elegant but risky. Huang pushed it anyway, and within months the team produced Wu Dao 1.0, then Wu Dao 2.0, at a time when most Chinese players had not yet committed to the paradigm.
The wager paid off in a specific, citable way. Emu3 discards diffusion and compositional vision-language stacks entirely. It tokenizes text, images and video into one shared discrete space and trains a single Transformer to predict the next token. The Nature reviewers noted that Emu3 matched dedicated task-specific models on both generation and perception, while also demonstrating coherent video generation, interleaved vision-language output, and vision-language-action control for robots.
From digital to physical
Huang's longer argument is that the next frontier is not bigger chatbots but models that understand the physical world. BAAI's 2025 "悟界 (Wu Jie)" release targeted exactly this: world models, cross-embodiment "brains," a digital-twin heart, and brain-science foundation models. In 2026 the institute followed with 悟界·Physis-v0.1, described as a domestic general physical-world base model.
A recurring thread in Huang's work is speed of observation. Borrowing from principles of the primate retina, BAAI has pursued "spiking vision" cameras that measure light intensity per pixel and compute only where the scene changes. The claim is that such sensors can feed world models high-fidelity, high-speed physical data that ordinary frame cameras cannot.
The practical payoff is continuity. Because Emu3 already expresses images, video and actions as the same kind of token, a robot task becomes just another next-token step rather than a separate vision stack bolted onto a language model. That is the bridge from a chatbot that describes the world to an agent that can act in it — and it is why Huang keeps returning to unified representation as the foundation, not a finishing touch.
Why it matters outside the lab
A Nature paper does not ship a product, but it signals where a research agenda is pointing. BAAI's stance — keep the unified, scalable route rather than chase the fastest demo — positions the institute as one of the few places treating embodied and world-model intelligence as a first-principles research problem rather than an engineering add-on. For China's AI field, the Emu3 result also matters symbolically: a domestically led large-model paper in the discipline's most cited journal.
Honest limitations
This article treats Emu3's Nature publication and the "first Chinese-led large-model paper in Nature" framing as reported by the journal and Chinese state science outlets; it does not re-run the experiments. Performance comparisons to diffusion or compositional baselines are summarized from the paper's abstract and secondary reporting, not independently benchmarked. BAAI's world-model and spiking-vision claims describe early-stage research; real-world deployment and commercial impact remain unproven. Open-source model counts are taken from institute communications and change continuously.
What readers can do now
- If you build with open models, BAAI's Emu3 and the Wu Jie series are worth testing for unified text-image-video tasks where a single framework simplifies your pipeline.
- For researchers, the next-token-prediction route is now a legitimate alternative to diffusion for multimodal work — read the Nature paper before assuming diffusion is mandatory.
- For non-specialists, the signal is that China's AI research is not only deploying models but also contesting the core methodological questions, and a Beijing institute just put that claim in a journal reviewers elsewhere respect.