A research lab in Beijing made a bet that most of the industry had quietly written off.
Instead of bolting image and video tools onto a language model, the team tried to do everything with one plain trick.
That bet just carried a Chinese model onto the pages of one of the world's top science journals.
One trick, three media
The model is Emu3 (悟界·Emu3), built by the Beijing Academy of Artificial Intelligence (北京智源人工智能研究院), known as BAAI (智源), under director Wang Zhongyuan (王仲远).
Its idea is stubbornly simple. Most multimodal systems stack pieces: a CLIP encoder for images, a diffusion model for generation, a language model for text. Emu3 throws that stack away.
- It turns images, text, and video into the same discrete tokens.
- One decoder-only transformer predicts the next token, no matter the type.
- There is no diffusion step and no separate perception pipeline — just next-token prediction (预测下一个词元).
The lab calls this native multimodality, and the claim is that a single objective can both understand (read a chart) and generate (draw a chart) without switching engines.
From a lab to Nature
Emu3 was first released in October 2024. The deeper result — that the next-token route scales to strong multimodal performance — was published as a peer-reviewed paper.
- Title: "Multimodal learning with next-token prediction for large multimodal models."
- Venue: Nature main journal, article
s41586-025-10041-x, posted online January 28–29, 2026, in print February 12, 2026. - Chinese state science outlets describe it as the first time a Chinese research institution led a large-model result into a Nature main-journal paper.
Nature's own editors noted that Emu3 learns text, image, and video together from one rule, and that the approach matters for building scalable, unified multimodal systems.
The base model carries about 13B parameters (130亿参数), modest by frontier standards, which is partly the point: the architecture, not the size, is the headline.
The bigger model: Emu3.5
BAAI did not stop at the paper. In late October 2025 it shipped Emu3.5, a clear step up.
- Scale jumped to roughly 34.1B parameters (341亿参数), trained on more than 13 trillion tokens of interleaved vision-language data.
- Context window reaches 32,768 tokens, long enough to reason over extended video.
- A method called DiDA (离散扩散自适应, Discrete Diffusion Adaptation) generates images in parallel instead of one token at a time, cutting per-image time by roughly 20x in the lab's tests.
On image tasks, BAAI reports Emu3.5 beating Gemini 2.5 Flash Image (Nano Banana) in head-to-head preference on world-exploration (65.5%) and embodied-manipulation (67.1%) prompts. Those are vendor-reported numbers with no independent leaderboard yet, so treat them as claims, not consensus.
Emu3's code is Apache-2.0, and the weights sit on Hugging Face, so the work is actually runnable, not just described.
Why the "world model" label
The interesting leap is in the framing. Emu3.5 is pitched as a world model (世界模型): trained on long internet video, it is meant to learn how scenes evolve — predict what happens next when something moves.
- That points at embodied intelligence (具身智能): a robot that can imagine the outcome of an action before taking it.
- BAAI ties this to its broader "悟界 (Wujie)" series, which also spans brain-science and life-science models.
Whether next-token prediction alone gets you a true world model is the open scientific question. The institute's answer is a scalable bet, not a finished product.
Who pays for this
BAAI is a non-profit institute, not a listed company, but it sits inside a heavy public push.
Beijing runs a municipal AI industry investment fund of RMB 20 billion (about US$2.8 billion / HK$21.8 billion), part of a wider plan to push the city's AI core-industry scale past RMB 1 trillion. Institutes like BAAI are exactly the kind of "do the hard, unprofitable research" actor that fund is meant to support.
That context explains the strategy: publish openly, open-source heavily, and let startups and universities carry the commercial load.
What readers can do now
- Read the Nature paper, not just the press: the method ablation is the real contribution, and it is freely available at
s41586-025-10041-x. - Pull Emu3 weights and test the "one objective" claim on your own image-and-text mixes; the Apache-2.0 license makes that legal and easy.
- Track Emu3.5 independently: because its headline numbers are vendor-reported, wait for third-party benchmarks before trusting them in production.
Honest limitations
- Emu3.5's strongest comparisons are self-reported; no independent replication was found in our review.
- "First Nature main-journal large-model paper from a Chinese institution" is a claim sourced from Chinese state media, not a verified global registry.
- A 34B model, even at 20x faster image generation, is slower per image than mature diffusion tools in practice; the released checkpoints remain pure next-token and may be slow.
- The world-model ambition is a research direction, not a delivered product, and next-token prediction alone may not capture physical causality.