BCI ReportFoundation modelEvaluated
LaBraM
5.82M parameters
A fixed configuration has been evaluated. See each task for channels and training mode.
Methods
EEG decoding methods with results on BCI Report
Every published figure for each method, by dataset.
The model register: every EEG model and paper we track, and how far we checked it (62 rows) →
Methods with pages
| Method | Family | Measured on |
|---|---|---|
| EEGNet | Compact model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 · Wearable SSVEP BCI dataset (dry and wet electrodes) · OpenBMI motor imagery (Lee et al. 2019) · BOAS · Bitbrain Open Access Sleep dataset |
| LaBraM | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 · Mobile BCI dataset (SSVEP and ERP paradigms) |
| CBraMod | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 · Mobile BCI dataset (SSVEP and ERP paradigms) · OpenBMI motor imagery (Lee et al. 2019) · BOAS · Bitbrain Open Access Sleep dataset · WBCIC-SHU motor imagery dataset |
| ShallowFBCSPNet | Compact model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| Deep4Net | Compact model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| CSP+LDA | Classical method | ds003810 · EEGMAT · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| CCA | Classical method | BETA · Wearable SSVEP BCI dataset (dry and wet electrodes) · Mobile BCI dataset (SSVEP and ERP paradigms) |
| FBCCA | Classical method | BETA · Wearable SSVEP BCI dataset (dry and wet electrodes) |
| eTRCA | — | Wearable SSVEP BCI dataset (dry and wet electrodes) |
| TS+LR and MDM | Classical method | ds003810 · EEGMAT · ds005342 |
| xDAWN+LDA and XDAWNCov+TS+LR | Classical method | ds006593 · TMNRED / ds005383 |
| REVE | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| LUNA | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| BrainOmni | Foundation model | ds003810 · EEGMAT · BETA · EESM19 scalp subset · ds005342 |
| CodeBrain | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| EEGMamba | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| ST-EEGFormer | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| eeg-fm-masking | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| ERP-FM | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| SingLEM | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| ZUNA | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
| EEGPT | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 |
Other measured baselines
BCI Report's own fixed baselines and references. Their figures sit in the topic pages and the core matrix.
- Spectral ridge · Classical method
- Temporal ridge · Classical method
BCI Report’s own comparator
Log band-power ridge (5 bands, visible channels)
Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds.
Beside the Spectral ridge row on seven core protocols: Motor imagery & rest, SSVEP · 8 channels, SSVEP · 4 channels, Arithmetic & rest, P300 target ERP, Semantic target ERP, Sleep staging.
How it differs from the Spectral ridge row
| Aspect | Spectral ridge | Log band-power ridge |
|---|---|---|
| Resolution | 90 bins of 0.5 Hz per channel (2-second and 30-second windows) or 45 bins of 1 Hz (1-second windows) | 5 broad bands per channel |
| Normalisation | each channel’s 0.5–45 Hz power sums to 1 | one divisor per window, the median channel standard deviation, for all channels |
| Clipping | none | ±15 median-channel standard deviations, before the spectrum |
| Estimator | Welch with constant detrend: one Hann segment on 2-second and 1-second windows; 29 half-overlapping 2-second segments averaged on 30-second windows | one periodic-Hann periodogram over the whole window |
| Window mean | removed per Welch segment | removed per window, before the input view |
| Sampling rate | 250 Hz | 200 Hz |
| Range | 0.5–45 Hz inclusive | 1–45 Hz, right-open |
| Transform | log of the relative density per bin | log of the absolute band integral, in scaled units |
| Head | the same scaler, alpha 100, solver, class weighting and folds | the same |
On BETA the Spectral ridge row is SSVEP-specific: log1p of the Hann-windowed projection powers at the 40 target frequencies and 3 harmonics, 120 features per channel. This comparator keeps the same 5 broad bands, 5 features per channel, as a control stated before scoring.
- Alpha is fixed at 100, as for every published ridge row; it was not tuned for 5 features per channel. Choosing the ridge alpha by the project's train-only rule (from 1, 100 and 10,000) in place of the fixed alpha 100 gave paired differences whose 95% intervals included zero on all seven protocols.
- Computing the same 5 bands on the microvolt windows (250 Hz, no per-window scaling or clipping) in place of the scaled 200 Hz view gave paired differences whose 95% intervals included zero on six protocols; on the 4-channel BETA protocol the difference, scaled view minus microvolt view, was +0.39 pp (95% interval +0.06 to +0.74), a small difference.
- On sleep staging the 30-second spectrum is one long periodogram; the Spectral ridge row averages 29 Welch segments of 2 seconds, and the frozen encoders average 15 segments of 2 seconds.
- The project’s own code, run with numpy, scipy and scikit-learn (BSD-3-Clause); no third-party weights or code were run. Each protocol’s data keep their released record.
Scored beside the matrix or in a panel of their own
Each links the section that holds its scores, with its comparator beside it.
- Signal-JEPA · Foundation model · frozen, with a ridge head, on motor imagery & rest and arithmetic & rest →
- EEGSimpleConv · Compact model · trained from scratch on the same two protocols →
- NeuroLM-B · Foundation model · frozen features on arithmetic & rest, and on an infant case-control cohort →
- NeuroRVQ · Foundation model · frozen backbone on three tasks →
- EEG-Captioner · Foundation model · frozen encoder on ten-second arithmetic & rest windows →
- AXON · Foundation model · frozen, on the same ten-second windows →
- LaBraM at Scale · Foundation model · frozen encoders on every core protocol →
In the directory, not in the matrix
Models listed in the model directory with no result in the core matrix or beside it, and the reason.
- BIOT ↗ · Foundation model · Adapter needed
- BENDR ↗ · Foundation model · Research candidate
- CSBrain ↗ · Foundation model · Access unverified
- Neuro-GPT ↗ · Foundation model · Research candidate
- EEG-Conformer ↗ · Compact model · Research candidate
- MIRepNet · Foundation model · Catalogue only (not evaluated)
- EEG-DINO · Foundation model · Catalogue only (not evaluated)
Catalogue
Model directory: from compact baselines to foundation models
Parameter counts depend on the backbone and task configuration.
Foundation modelEvaluated
EEGPT
25.29M parameters
Evaluated locally, but no score is published in this release: the checkpoint license is unresolved. Results are withheld pending that review, not because the model failed to run.
Released status: Rights review pending. Evaluated as a frozen encoder on the eight core protocols, in the same runs as the published LaBraM rows and, except on idle, the CBraMod rows, and in one pilot comparison outside every leaderboard. The checkpoint’s licence remains unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. Its core-matrix cells and pretraining-attribution rows are published on its page; the cells also sit beside the core matrix on every protocol page. The note above is the released one. The pilot comparison →
Foundation modelEvaluated
CBraMod
4.88M parameters
A fixed configuration has been evaluated. See each task for channels and training mode.
Compact modelEvaluated
EEGNet
1.9K parameters
A fixed configuration has been evaluated. See each task for channels and training mode.
Compact modelEvaluated
ShallowFBCSPNet
30.5K parameters
A fixed configuration has been evaluated. See each task for channels and training mode.
Compact modelEvaluated
Deep4Net
274.6K parameters
A fixed configuration has been evaluated. See each task for channels and training mode.
Classical methodEvaluated
CSP+LDA
— parameters
A fixed configuration has been evaluated. See each task for channels and training mode.
Foundation modelAdapter needed
BIOT
3.19M parameters
Pretrained bipolar tokens do not directly match the unipolar recordings. A separately validated adapter is required.
Foundation modelEvaluated
REVE Base
69.19M parameters
Code and electrode positions are accessible. Base weights require accepting the official access agreement; they have not been downloaded.
Released status: Access gated. The owner accepted the REVE Responsible Use License v1.0, and the foundation-model evaluation ran the base weights as a frozen probe on the eight core protocols and adapted them on EEGMAT. Aggregate scientific results only; no adapted weights are shared. The note above is the released one. Source ↗
Weights: REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared.
Foundation modelEvaluated (two protocols, beside the matrix)
Signal-JEPA
3.46M parameters
Public, ungated weights; research candidate only; not downloaded or executed for this release. The official model card provides a 13,847,488-byte safetensors checkpoint trained on Lee2019 at 128 Hz. All 17 ds005342 channel names match its 62-channel pretraining layout, but a separate locked adaptation protocol is still required.
Released status: Research candidate. Scored as a frozen encoder with a ridge head on motor imagery & rest and arithmetic & rest, beside the core matrix, with its own input adapter (0.5–40 Hz, 128 Hz). Aggregate results only; the weights are not redistributed. The note above is the released one. Its rows on the two protocol pages →
Foundation modelResearch candidate
BENDR
157.00M parameters
Public, ungated weights; research candidate only; not downloaded or executed in this expansion. The Braindecode checkpoint is 628,580,476 bytes and expects 20 channels at 250 Hz. A documented montage or input adapter is required for 17-channel data.
Foundation modelAccess unverified
CSBrain
— parameters
Public source and an official Google Drive weight link are listed; anonymous weight download, file size, and execution remain unverified. The backbone accepts caller-supplied brain regions and channel ordering, while the published BCIC head is fixed to 22 channels. A new adapter and a checkpoint-key loading audit are required.
Foundation modelResearch candidate
Neuro-GPT
— parameters
Public code and Hugging Face weights; research candidate only; not downloaded or executed for this release. The model page shows an approximately 318 MB checkpoint. The published EEGConformer input path is fixed to 22 channels, so using 17 channels requires an explicit mapping or replacement of the spatial layer.
Compact modelResearch candidate
EEG-Conformer
— parameters
Public source; no official pretrained checkpoint was found; not executed for this release. This is a from-scratch baseline rather than a foundation-model checkpoint. The author script fixes the spatial convolution to 22 channels and must be parameterized for other montages.
Compact modelEvaluated (two protocols, beside the matrix)
EEGSimpleConv
— parameters
Public implementation; no pretrained weights; not executed for this release. The maintained Braindecode implementation accepts the channel count and sampling rate directly, making it a practical lightweight from-scratch comparator.
Released status: Research candidate. Trained from scratch for 20 epochs on motor imagery & rest and arithmetic & rest, three seeds, beside the core matrix; the pages print the mean of the three seeds. The note above is the released one. Its rows on the two protocol pages →
Classical methodEvaluated
Standard CCA
— parameters
Fixed reference method; inspect each task for input features and fit protocol.
Classical methodEvaluated
Spectral ridge
— parameters
Fixed reference method; inspect each task for input features and fit protocol.
Classical methodEvaluated
Temporal ridge
— parameters
Fixed reference method; inspect each task for input features and fit protocol.
Foundation modelEvaluated
REVE Large
389.98M parameters
Frozen probes on all eight core protocols and the EEGMAT adaptation, with the same corpus and terms as REVE Base.
Weights: REVE Responsible Use License v1.0 (as REVE Base).
Foundation modelEvaluated
LUNA (Base, Large)
6.57M · 40.45M parameters
Base · Large
Frozen probes on all eight core protocols; Base was also adapted on EEGMAT, Large was not, by design.
Weights: CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release.
Foundation modelEvaluated (6 of 8 protocols)
BrainOmni Base
31.84M parameters
Frozen probes on six of the eight core protocols and the EEGMAT adaptation. The two one-second ERP protocols were not run: its tokenizer needs two-second windows.
Foundation modelEvaluated
CodeBrain
13.53M parameters
Frozen probes on all eight core protocols and the EEGMAT adaptation, pooled as the published CBraMod rows are.
Foundation modelEvaluated
EEGMamba
3.28M parameters
The Neural Networks (2025) model from the CBraMod group. Frozen probes on all eight core protocols and the EEGMAT adaptation.
Foundation modelEvaluated
ST-EEGFormer (Base, Large)
85.18M · 303.00M parameters
Base · Large
Frozen probes on all eight core protocols; Base was also adapted on EEGMAT. BETA is in the authors’ published pretraining list, so both BETA protocols are flagged.
Weights: MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name).
Foundation modelEvaluated (frozen probes, masking ablation)
eeg-fm-masking (4 checkpoints)
12.69M parameters
Four checkpoints of one masking study, as frozen probes on all eight core protocols: the paper-recommended MAE checkpoint as a matrix row, three siblings in a masking ablation.
Weights: CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487).
Foundation modelEvaluated (frozen probes)
ERP-FM Base
2.02M parameters
Frozen probes on all eight core protocols: the two ERP protocols are in its design, the other six are negative controls.
Weights: CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed.
Foundation modelEvaluated
SingLEM
3.27M parameters
A single-channel encoder. Frozen probes on all eight core protocols and the EEGMAT adaptation. TMNRED is in the authors’ published pretraining list, so the semantic-target protocol is flagged.
Weights: MIT.
Foundation modelEvaluated
ZUNA 1.1
172.10M parameters
Frozen probes on all eight core protocols and the EEGMAT adaptation. Its authors publish no pretraining list, so every cell is marked exposure unknown.
Weights: Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use.
Foundation modelCatalogue only (not evaluated)
MIRepNet
— parameters
Listed only, by the owner’s decision.
Foundation modelCatalogue only (not evaluated)
EEG-DINO
— parameters
Listed only, by the owner’s decision. The upstream weights carry no licence.
External
Field notes: other people’s work
Curated research updates, linked to original sources.
Benchmark
A new benchmark compares EEG foundation models with task-specific decoders ↗
Researchers surveyed 55 representative models and evaluated 12 open-source foundation models with task-specific baselines across 13 datasets and nine BCI paradigms. The study reports that linear probing is often insufficient and that larger models do not automatically generalize better, reinforcing the need for fixed cross-subject and few-shot calibration protocols.
Original arXiv paperModel release
REVE releases an EEG foundation model for varying electrode layouts ↗
REVE reports pretraining on 92 datasets, about 60,000 hours of EEG, and 25,000 participants, with position encodings designed for different electrode arrangements. Its public code, weights, and tutorials make cross-montage transfer an auditable evaluation target. Access to the Base checkpoint is gated; it is not yet available in our local test pool.
REVE Base and REVE Large have since been evaluated here; their frozen-probe rows are on every protocol page. REVE’s method page →
Original arXiv paper and NeurIPS 2025 paper pageResearch release
NeurIPT brings heterogeneous EEG configurations into a unified pretraining framework ↗
NeurIPT models homogeneous and heterogeneous spatiotemporal relationships across BCI tasks and electrode layouts and appeared as a NeurIPS 2025 poster. Its official repository still says that code restoration is in progress, so it is a research update rather than a reproducible model card.
OpenReview / NeurIPS 2025Model release
CSBrain adds cross-scale spatiotemporal structure to general EEG decoding ↗
CSBrain combines local temporal windows, brain regions, and structured sparse attention, with reported experiments spanning 16 datasets and 11 task types. The official repository provides pretraining and downstream scripts plus a public weight entry point, making it a candidate for a later unified-protocol evaluation.
Original arXiv paper and NeurIPS 2025 paper pageReal-time BCI
An eight-channel, low-preprocessing self-supervised EEG model targets real-time BCI ↗
This study applies HuBERT-style self-supervised learning to eight-channel scalp EEG and examines motor imagery, P300, subject differences, and alpha rhythms. It motivates a separate deployment-oriented track for low channel counts, limited preprocessing, and runtime budgets.
Original arXiv paper