A sleep lab has spent years recording the same thing in incompatible ways. One site collects EEG from a 64-channel cap. Another has magnetoencephalography in a shielded room. A third has calcium imaging from mice and Neuropixels recordings from macaques. None of it talks to the rest. Every new question means a new bespoke pipeline, written once, used once, thrown away.
That is the mess two Chinese labs are trying to tokenise. In 2026, the Beijing Academy of Artificial Intelligence (北京智源人工智能研究院) released Brainμ 1.0, and the Shanghai Artificial Intelligence Laboratory shipped BrainOmni inside its broader science platform. Both make the same bet: neural signals, whatever the machine that produced them, can be encoded into a shared representation and trained on like language.
Key takeaways
- Brainμ 1.0 unifies signals across humans, macaques and mice, and across EEG, fMRI, MEG, fNIRS, Neuropixels and calcium imaging, by encoding them as standard tokens aligned with text, image and video.
- BrainOmni claims to be the first brain foundation model to unify EEG and MEG in one architecture, with results published in a Cell sub-journal and third-party benchmarking described as best-in-class.
- The point is not a chatbot for the brain. It is to remove the per-study engineering tax that makes every brain-computer interface (脑机接口) decoder start from scratch.
- Both models are research infrastructure. Neither is a medical device, and neither has been validated as a diagnostic.
What a neural foundation model actually is
A language model learns that "the" predicts a noun. A neural foundation model learns what a waveform predicts — that a certain pattern in motor cortex precedes a hand movement, that a spindle in a sleep recording means a particular stage.
The trick that makes this possible is a tokenizer. Brainμ trains a "Brain Tokenizer" that converts raw neural signal into discrete tokens, then aligns those tokens against the token space of a pre-trained large model (大模型). Once brain activity and language live in the same space, a decoder can be fine-tuned rather than built.
Practically, this changes the unit of work. Instead of "collect 40 hours from this patient, train a decoder for this patient," the pitch is "pre-train once on everything, adapt per patient with a few minutes of data" — the difference that decides whether a device can be used on day one or day thirty.
Brainμ: three species, one token space
BAAI's description is specific. Brainμ 1.0 covers three species — human, macaque, mouse — and takes in EEG, fMRI, MEG, fNIRS, Neuropixels and calcium imaging. It encodes all of them into a single token format, then aligns brain tokens with language, images and video so a pre-trained model's knowledge can be reused.
Alongside it sits BrainToken, described as the largest AI-ready neuroscience data platform, with over a trillion tokens of neural data after full tokenisation — roughly an order of magnitude more neural data than previous mainstream base models, according to the institute. The institute says it is working with 32 research institutions and hospitals, including a collaboration with Xuanwu Hospital on personalised early screening for Alzheimer's disease.
Two results carry the claim. A joint study with Tsinghua University on how memory regulates sleep was published in Science, with Brainμ handling the multimodal analysis; the reported finding is that memory and sleep regulate each other, with positive memory content improving sleep quality and negative content worsening fragmentation. Separately, BAAI reports the model's training covered more than 70,000 nights of sleep recordings and has powered automated analysis in partner labs for over twelve consecutive months.
BrainOmni: the EEG–MEG merge
Shanghai's version attacks a narrower, messier problem. EEG measures electrical potential at the scalp; MEG measures the magnetic field. Devices differ, channel layouts differ, and subjects differ — which is why most models are built for one dataset and die there.
BrainOmni is described by the Shanghai Artificial Intelligence Laboratory as the first foundation model to unify EEG and MEG signals, modelling across devices, modalities, tasks and subjects. Its results were published in a Cell sub-journal; the lab says third-party evaluation ranks it as the strongest EEG foundation model overall, and that it leads on tasks including suicide-risk detection from speech and psychological crisis screening. It also reports inbound requests from universities in Europe and the US and from MEG equipment makers.
For anyone building wearables, this is the more immediately relevant of the two: consumer EEG headbands are cheap, noisy and wildly inconsistent between users, which is exactly the regime a cross-subject model is supposed to survive.
Why brain-computer interface people should care
Decoders today are patient-specific and session-specific. Signal drifts as electrodes scar, as attention wanders, as the user gets tired, and the model degrades until someone recalibrates it. A model pre-trained across thousands of subjects starts each new user from a better place, and — more importantly — can be updated centrally rather than rebuilt per clinic.
That is an industrial argument, not an academic one. A brain-computer interface (脑机接口) company shipping to many hospitals needs software that improves for all of them at once.
According to Li An, chief scientist at BrainNet (脑机网), the arrival of neural foundation models shifts the competitive axis in brain-computer interfaces from hardware channel counts to who owns the largest clean, aligned, longitudinal dataset.
Honest limitations
Both models are described primarily by their own institutions, though the descriptions were carried by state broadcasters and national newspapers rather than only by press releases. "World's first" and "best overall" are institutional claims; I have not found an independent benchmark reproducing them, and the underlying Cell sub-journal and Science papers are cited here only through institutional summaries, not read directly. Publication in a journal is not validation of clinical usefulness: neither model is a registered medical device, neither has prospective clinical trial (临床试验) evidence for diagnosis, and claims about Alzheimer's screening and suicide-risk detection are early. Cross-species and cross-modality alignment can fail quietly on populations unlike the training set. This piece does not address data consent and neural privacy, the compute cost of training, or how much of the underlying data is actually accessible to outside researchers.
What readers can do now
- If you work with neural data, start treating metadata as the asset: channel layout, sampling rate, device model and task labels are what make a signal reusable. Most datasets fail on this, not on signal quality.
- Clinicians evaluating "AI for brain signals" should ask one question: was the model tested on subjects and devices it never saw during training? Cross-subject performance is the only number that predicts bedside behaviour.
- Patients should assume that any model claiming to screen for depression or Alzheimer's from brain signals is a research tool until a regulator says otherwise.