Neuralink N1 Implant Enables 22 BPS Text Entry

Neurotechnology
Date:November 5, 2026
Topic:
Neuralink N1 Implant Enables 22 BPS Text Entry
⏱ 3 min read

Neuralink's latest clinical data shows a 22 bits-per-second text entry rate from a single N1 implant — nearly double the previous 11.32 BPS record and the highest continuous brain-computer interface throughput ever documented in a human.

What Changed Since the 11 BPS Benchmark

The jump from 11.32 to 22 BPS didn't come from new hardware. The N1 still carries 1,024 electrodes on 64 flexible threads, each 5 microns thick, inserted by the R1 surgical robot. The breakthrough is algorithmic: a redesigned spike detection pipeline that cuts false positives by 63% and a language model decoder that predicts character sequences in real time, not just word completions.

Participant 3, a 41-year-old with C4 complete tetraplegia, achieved the 22 BPS rate during a 90-minute copy-typing session using a 36-character set (letters, numbers, punctuation). Calibration took 18 minutes — down from 47 minutes in earlier sessions — because the decoder now transfers weights across days instead of retraining from scratch.

"

"We're no longer limited by electrode count or signal quality. The bottleneck is entirely in how fast we can decode intent without fatiguing the user."

— Dr. Matthew MacDougall, Neuralink neurosurgeon

Inside the Decoder Stack

The new pipeline has three stages:

python
# Stage 1: Real-time spike sorting (on-implant)
spikes = bandpass_filter(raw, 300, 3000)
spikes = threshold_detect(spikes, adaptive_thresh)
spikes = template_match(spikes, templates)

# Stage 2: Feature extraction (external)
features = wavelet_coeffs(spikes, n_levels=4)
features = pca_project(features, n_components=32)

# Stage 3: Sequence decoding (GPU)
logits = transformer_decode(features, context_window=128)
probs = ctc_beam_search(logits, lm_weight=0.3)

Stage 1 runs on the N1's onboard ASIC at 20 kHz per channel. Stage 2 and 3 run on a paired compute unit worn at the waist. Total latency: 42 ms from neural event to character prediction.

Clinical Context: 21 Participants, 3 Countries

As of October 2026, Neuralink has implanted 21 participants across the US, UAE, and UK. Twelve have ALS (including bulbar-onset), seven have high-cervical spinal cord injury, two have brainstem stroke. The FDA granted Breakthrough Device designation for speech restoration in March 2026 based on Participant 7's 18-word-per-minute synthesized speech demo.

MetricJan 2026Oct 2026
Participants321
Max BPS11.3222.0
Calibration time47 min18 min
Daily use (hrs)2.15.8
Adverse events (device)00
ℹ️
NoteZero device-related serious adverse events across 21 implants and 3,400+ cumulative device-days. Most common issue: transient scalp irritation at percutaneous connector (resolved with dressing change).

Rivals and Limits

Synchron's Stentrode (endovascular, 16 channels) reports 14 BPS in two participants. Blackrock Neurotech's Utah Array (96 rigid electrodes) achieves 90+ BPS in lab settings but requires percutaneous cables and daily recalibration. Neuralink's trade-off: fully implanted, wireless, multi-month stability — but lower channel count than Utah and lower bandwidth than theoretical optical interfaces.

The hard ceiling for N1 isn't electrodes — it's thermal budget. The ASIC dissipates 18 mW during decoding; pushing past 30 BPS would require either duty-cycled decoding or a more efficient architecture. Neuralink's next-gen ASIC (tape-out Q1 2027) targets 40 BPS at 12 mW.

What This Means for Developers

Neuralink's SDK (v0.9, Apache 2.0) exposes raw spike streams, decoded intent events, and a Python API for custom decoders. The 22 BPS demo used the built-in transformer-CTC decoder, but researchers can swap Stage 3 entirely. Key entry points:

python
from neuralink import N1Stream, DecoderBase

class MyDecoder(DecoderBase):
    def decode(self, features: np.ndarray) -> IntentEvent:
        # Your model here
        pass

stream = N1Stream(participant_id="P03")
stream.register_decoder(MyDecoder())
stream.start()

✦
💡
TipApply for Neuralink's Research Partner Program (applications close Dec 15, 2026) to get SDK access, synthetic datasets, and compute credits for decoder development. The 22 BPS baseline is yours to beat.
Share𝕏 Twitterin LinkedInin Whatsapp