A new EEG model trained on 25,000 hours of brain recordings can read 100x longer signals at 55% of the memory

Continuous brain-wave monitoring, the kind used in epilepsy wards, intensive care units, and brain-computer interfaces, runs into a stubborn wall long before the patient does. Every extra hour of recording multiplies the memory that the standard Transformer needs to keep its attention sharp. A 24-hour EEG session can be the difference between a model that fits on one machine and a model that does not fit at all.

S-CEReBrO, a foundation model accepted to MICCAI 2026 and posted to arXiv this week, breaks that wall with a reworked attention scheme that holds memory constant no matter how long the recording grows. The model was trained on more than 25,000 hours of EEG from over 12,000 subjects.

The mechanism is called Windowed Alternating Attention. Instead of letting every token in a multi-hour stream attend to every other token (the default Transformer behavior that exhausts GPU memory on long sequences), the model alternates between short-range windows in time and short-range windows in space. The result is a KV cache, the working memory the model uses during inference, whose size no longer scales with recording length.

On the authors' own benchmarks, S-CEReBrO can process signals 100 times longer than full self-attention, 3 times longer than low-rank linear attention (a faster but cruder alternative), and uses 55% of the memory that low-rank linear attention needs on long contexts, with 2.1 times the inference throughput. On a downstream benchmark suite of 11 EEG tasks, the model reaches state-of-the-art on 7, with up to 60% fewer parameters than competing approaches.

Companion code accompanies the paper. The 7-of-11 result is a benchmark ceiling, not a clinical deployment. The paper is a preprint, accepted at MICCAI 2026 but not yet at camera-ready, and not yet independently reproduced. The 25,000-hour pretraining corpus and the parameter-efficiency number are evidence the architecture is doing real work at scale, not a signal that a wearable seizure detector is shipping next quarter.

The advance lands in one of the most demanding long-sequence settings in AI. The same memory wall that complicates million-token language models and hour-long video models also limits how long any single brain signal can be, and continuous EEG is one of the harder cases. A 72-hour ICU stream is a routine input. For population-scale neuroscience, where studies pull together tens of thousands of recordings from different sites, a foundation model that can ingest a full overnight session without splitting it into chunks lowers the engineering tax on the field.

Whether the architecture generalizes from the preprint's benchmark suite to a bedside monitor is the next test, and the camera-ready MICCAI version is where that work begins.