Sabi is building a consumer brain-computer interface that looks like a beanie and, according to the company, can turn thoughts into text.

The San Francisco startup emerged from three years in stealth with plans to launch its first product before the end of 2026. Sabi says the wearable will contain up to 100,000 miniature EEG sensors and use machine learning to decode intended speech at around 30 words per minute.

If those numbers hold up, it would represent a major jump for non-invasive brain-computer interfaces.

Inside the beanie

Sabi’s approach starts with sensor density. Conventional EEG systems use 32 to 64 electrodes placed across the scalp to detect electrical activity generated by the brain. Even advanced research systems generally operate with hundreds of channels.

Sabi says it can embed as many as 100,000 miniature sensors throughout the fabric of a beanie. The company argues that collecting signals across so many locations makes it easier to estimate where neural activity originates and gives its decoding models richer data to work with.

That hardware is  being designed as an everyday wearable. Sabi says the beanie will be breathable, washable, and usable across different head shapes, hairstyles, and environments. A baseball-cap version is also planned.

Those practical details spark interest. High-quality EEG usually requires careful electrode placement and reliable contact with the scalp. Delivering that inside clothing people can wear all day creates a host of  engineering problems.

Turning thoughts into text

Sabi’s main goal is communication. The company wants users to compose text through internal speech, effectively replacing some of the input normally handled by a keyboard.

It claims to have already achieved around 30 words per minute with what it describes as reasonable accuracy. That would still sit well below natural speech, which commonly exceeds 140 words per minute, but it would be unusually fast for a non-invasive EEG system.

For comparison, some of the strongest results in speech BCIs have come from electrodes implanted directly in the brain. A Stanford study of inner-speech decoding reached substantially higher communication rates using intracortical recordings, where the signal is much richer than anything measured through the scalp.

Sabi also says its models can identify the intention to communicate, allowing the system to distinguish deliberate input from unrelated thoughts. That would address one of the central usability problems with any always-on thought-to-text interface.

Will the claims hold?

The concept is compelling because thought-to-text is one of the clearest possible consumer uses for a BCI. But Sabi wants to achieve that signal quality  through a wearable EEG system, which is known for high signal-to-noise ratios.

A jump from hundreds of EEG channels to 100,000 sensors would itself be significant. Combining that with reliable inner-speech decoding, everyday wearability, and 30-word-per-minute communication raises the bar considerably higher.

Sabi has not yet published enough technical data to independently assess those claims. Details around accuracy, vocabulary size, participant numbers, training requirements, sensor performance, and real-world testing will be central once the product approaches launch.

For a deeper look at Sabi’s technology and how its claims compare with current speech BCIs, read the full Neurofounders piece.