Foundation models are becoming central to developing brain-computer interfaces, and Neuralink has just offered a first look into how it approaches that challenge. In a recently published blog post, the firm says its trial participants have generated more than 50,000 hours of intracortical recordings. That dataset is being used to pretrain a model that learns from neural activity collected during everyday BCI use.

The foundation model is first aimed at a central BCI challenge. Neural signals change over time, meaning that decoders often need to be recalibrated to maintain performance. Neuralink says its pretrained models have reduced calibration from around ten minutes per day to ten minutes per week. Beyond calibration, six participants achieved their fastest-ever scores on Neuralink’s cursor-control benchmark.

Pretraining on 50,000 Hours of Brain Data

BCI decoders translate raw neural signals into functional commands, such as moving a cursor. They are typically trained through labelled calibration sessions, where neural activity is recorded alongside a known intended movement. But the signals picked up by individual electrodes shift over time, gradually reducing decoder performance and forcing users to repeat that calibration.

Calibration sessions represent only a small fraction of the neural data generated by a BCI. During everyday use, participants continuously produce recordings without a corresponding label for every movement or action. Neuralink’s first participant alone has generated more than 9,000 hours of this largely unlabelled data.

Neuralink is now using those recordings for self-supervised pretraining. The model learns recurring structure in neural activity without requiring every recording to be paired with a known movement, creating representations that support downstream cursor decoders. 

Neuralink says it has applied the approach across more than 50,000 hours of recordings, using participant-specific models trained on each user’s own data.

The initial gains mostly show up in stability. Neuralink says participants previously spent an average of around 55 minutes each week recalibrating their decoders. For some, pretraining has reduced that to just ten minutes per week, while some decoders have continued performing strongly for several weeks. 

The company also showed that a decoder calibrated 20 months earlier could still provide usable cursor control, suggesting that the learned representations are less sensitive to changes in the underlying neural signal.

Control performance also improved. Six participants set personal records on Neuralink’s Webgrid task, where users move a cursor between highlighted targets as quickly and accurately as possible. Performance is expressed in bits per second.

Neuralink says able-bodied users controlling a mouse typically score around 8-10 bits per second on the task. One of Neuralink’s participants, using its new model, reached 11.32.

Building Models on Real-World Use

Neuralink is rapidly expanding its clinical programme. The company implanted its first participant in January 2024 and reported 21 participants by January 2026. The implant is now used for much more than short research demos. One UK participant reportedly used Telepathy for as much as 17 hours in a day.

In September, the company also showed its work in speech restoration. Terry, a participant living with ALS in Neuralink’s VOICE study, used neural activity associated with intended speech to generate spoken output through a synthetic version of his voice.

Such real-world use is creating a data advantage. Unlike large language models, brain foundation models depend on neural recordings that must first be collected through physical interfaces. For implanted systems, that ties model development closely to clinical deployment. Neuralink’s first participant alone has contributed more than 9,000 hours to the company’s 50,000-hour dataset.

Read the full piece

Inside the Frontier of Brain Foundation Models

Brain foundation models aim to generalize EEG across people, devices and tasks. Arctop and Piramidal discuss scale, representation and real-world deployment.

Neuralink is not alone in its foundation model work. Synchron introduced Chiral in 2025, a planned foundation model trained on data from its Stentrode implants. More recently, Neurosoft announced a partnership with Mila on a cortical foundation model. Synaptrix places the model layer even closer to the center of its platform, combining its own neural datasets with brain-specific AI to train models across devices and recording contexts.

For a deeper look at where brain foundation models are heading, Neurofounders recently explored the field with researchers and founders working directly on these systems. Watch the full webinar here.