NeuroDX released the EEG model MANAS-1-small with 40 million parameters

Company NeuroDX introduced MANAS-1-small — a specialized foundational model for electroencephalogram (EEG) analysis. The architecture contains 40 million parameters and is optimized for working with neurophysiological data.

The model demonstrated state-of-the-art results within the EEG-FM-Bench benchmark. During testing, MANAS-1-small showed higher classification accuracy and signal processing compared to existing analogs, including specialized solutions REVE and CBraMod, on various brain activity analysis tasks.

Developers published the model weights in open access on the Hugging Face platform. The compact size of 40M parameters allows deploying the system on local computing resources, which is critical for medical applications requiring real-time data processing without transmitting information to external servers.

The emergence of specialized models of this scale lowers the entry barrier for researchers in the field of neurointerfaces and clinical diagnostics. Using the pre-trained weights of MANAS-1-small reduces the time required for fine-tuning for specific EEG analysis tasks, ensuring high pattern recognition accuracy compared to general-purpose models.