Descartes

The heart has an ECG; the lungs have spirometry. The brain has no equivalent: depression, PTSD, post-concussive syndrome and early cognitive decline are still diagnosed from questionnaires and behavioral observation, so two patients with the same label can have very different underlying physiology and no way to tell which treatment fits which brain. Scalp EEG is cheap, non-invasive and rich in signal, but the same cognitive state produces waveforms that look completely different from one person to the next, which has historically forced EEG decoders to be fitted per subject or per cohort.

Descartes is Hemispheric's response to that variability problem: a large model pretrained on brain activity, which the company says is applied as a single fixed checkpoint to people it was never trained on. Hemispheric, based in Tel Aviv, announced the model on 15 July 2026 alongside its emergence from six years of stealth with $52 million in early-stage funding. The company's argument is a scaling argument rather than an architectural one — that with enough recorded hours across enough people, inter-subject variability stops being noise to be normalized away and becomes structure a model can learn, in the way that scale turned text and image modeling into general-purpose representation learning.

That framing places Descartes alongside academic EEG foundation models such as LaBraM, CBraMod and BrainOmni, which pretrain on aggregated public recordings and release weights for the community. The difference is the corpus and the posture: Hemispheric's training data is proprietary and collected through its own global research network, and the company sells access to models trained on that data rather than the data itself. (The name is shared with an unrelated single-cell atlas of human gene expression during development; the two projects have nothing in common beyond the word.)

#Key Features

• Cross-subject decoding: Hemispheric states that Descartes is evaluated on thousands of participants who were not in its training set, and generalizes to decoding tasks it was not fitted on.
• Multimodal pretraining: the training corpus pairs EEG with behavioral responses collected during perceptual, cognitive and emotional tasks, rather than resting-state traces alone.
• Consumer-grade acquisition: a session uses a lightweight dry-electrode headset worn for roughly 15 minutes while the participant works through an app on a tablet or phone — no gel and no clinical EEG suite.
• Clinical readouts as the output: the company positions the model's outputs as objective measures a clinician can use for stratification, treatment selection and longitudinal monitoring, not as raw embeddings.
• Access by application only: Hemispheric offers Descartes through a hosted API that prospective users apply for; no code, weights or checkpoints have been released.

#Technical Details

Hemispheric describes Descartes 1.0 as a 6-billion-parameter model trained on more than 250,000 hours of multimodal EEG and behavioral recordings from over 100,000 participants, gathered over six years through purpose-built data collection laboratories by a 112-person team. The company also states that brain decoding follows the same scaling laws as language and vision — its central technical claim, and the reason the corpus is described in hours rather than in datasets.

Every one of those figures is a company statement. No preprint, technical report or peer-reviewed publication describing Descartes has been released, and the architecture, tokenization scheme, electrode montage, pretraining objective and evaluation protocol are not public. No benchmark numbers have been published against the standard EEG evaluation suites that academic brain foundation models report on, so the accuracy claims cannot be compared with published work.

#Applications

The first applications Hemispheric names are in precision brain health: PTSD, mild traumatic brain injury, depression, anxiety, schizophrenia and Alzheimer's disease, where the intended use is measuring severity, separating a diagnosis into biological subtypes, and choosing among treatments that today are selected by trial and error. The company also positions the model as an instrumentation layer for CNS drug and device development — measuring and predicting treatment response and optimizing protocols — and is working with government and pharmaceutical partners toward US and European deployment. Hemispheric has presented the platform to leadership at the FDA's Center for Devices and Radiological Health and describes a regulatory strategy in progress; no clearance or approval has been granted.

#Impact

Descartes is a bet that the recipe which produced general-purpose language and vision models transfers to non-invasive electrophysiology, at a corpus scale academic EEG groups cannot match from public data. If the cross-subject generalization holds, it would matter beyond one company's product line, because the limiting factor in clinical EEG has always been that decoders do not transfer between people. That remains unestablished outside Hemispheric's own evaluation: with no paper, no weights and no code, the model cannot be reproduced or independently checked, and this entry records what the developer states rather than what has been demonstrated in public.