Model reference · open weights

reve

Embeddings brain-bzh Embeddings 1 build Its own licence terms 22k dl/mo

reve is an open-weight embedding model from brain-bzh. reve-base (FP32) weighs 138 MB; the smallest configuration that runs it is RTX 3060 12 GB.

REVE is a 69M parameter transformer-based foundation model developed by brain-bzh for EEG signal processing and feature extraction.
It was pretrained on over 60,000 hours of data from 92 datasets and uses a 4D positional encoding scheme to handle arbitrary signal lengths and electrode configurations.
The model is available under the REVE Responsible Use License v1.0.

Summary of the brain-bzh/reve-base model card, 2026-10-01

What it is

What it runs on

Memory and cards for reve-base (FP32)

138 MBweights, file size

1.1 GBruntime overhead

How it works

Benchmarks

As published on the model card: the maker's own numbers, not measured by AxForge.

Task	Dataset	Metric	Score
feature-extraction	TUAB	Accuracy	0.832
feature-extraction	TUEV	Accuracy	0.676
feature-extraction	PhysionetMI	Accuracy	0.648
feature-extraction	BCICIV2a	Accuracy	0.640
feature-extraction	FACED	Accuracy	0.565
feature-extraction	ISRUC	Accuracy	0.782
feature-extraction	Mumtaz	Accuracy	0.964
feature-extraction	MentalArithmetic	Accuracy	0.766
feature-extraction	BCI2020-3	Accuracy	0.564
feature-extraction	TUAB-LP	Accuracy	0.810
feature-extraction	TUEV-LP	Accuracy	0.592
feature-extraction	PhysionetMI-LP	Accuracy	0.537
feature-extraction	BCICIV2a-LP	Accuracy	0.517
feature-extraction	ISRUC-LP	Accuracy	0.697
feature-extraction	Mumtaz-LP	Accuracy	0.962
feature-extraction	MentalArithmetic-LP	Accuracy	0.740
feature-extraction	BCII2020-3-LP	Accuracy	0.390

Details