NDT3

Intracortical brain-computer interfaces read spiking activity from microelectrode arrays in motor cortex and translate it into cursor movement, robotic arm control, or attempted speech. Historically, progress came from restricting complexity: a decoder is fit to one participant, one array, one task, often one session, and is recalibrated whenever anything drifts. That works, but it means every new participant starts from zero and every session burns time on calibration.

NDT3 — Neural Data Transformer 3 — asks whether the foundation model recipe transfers to this domain. It comes from the University of Pittsburgh and Carnegie Mellon University, with data contributed by BCI labs at Northwestern University, the University of Chicago, the University of Kansas Medical Center, and Lawrence Berkeley National Laboratory, it pretrains an autoregressive transformer on 2,000 hours of neural population spiking activity paired with the motor covariates recorded alongside it, drawn from over 30 monkeys and human participants. The result is a single checkpoint that improves decoding on 8 downstream tasks and holds up under a variety of neural distribution shifts.

What distinguishes the paper from most foundation model releases is its candour about ceilings. The authors report that simply scaling autoregressive transformers seems unlikely to resolve the limitations that come from sensor variability and output stereotypy — the fact that electrode arrays differ unpredictably between implants, and that the behavioural outputs in these datasets are highly repetitive. Those are properties of the data, not of the architecture, and more parameters do not fix them.

#Key Features

• Paired pretraining objective: Rather than modeling spikes alone, NDT3 is trained autoregressively over interleaved neural activity and the motor covariates it accompanies, so behaviour is part of the sequence rather than a separate readout head.
• Cross-species corpus: Monkey and human intracortical recordings sit in one pretraining set, letting the abundant animal data support the scarce human clinical data.
• Demonstrated transfer: The same pretrained weights benefit 8 distinct downstream decoding tasks, and generalize across shifts in subject, session, and array configuration.
• Two released scales: A 45M-parameter model and a 350M-parameter model are both published, alongside checkpoints pretrained on 200-hour, 1,000-hour, and 2,000-hour subsets, which makes the data-scaling behaviour directly reproducible.
• Fine-tuning path included: The repository ships task-specific fine-tuned checkpoints and a worked notebook, so a new dataset is an adaptation problem rather than a from-scratch training run.

#Technical Details

NDT3 is an autoregressive transformer trained with bfloat16 precision and FlashAttention 2, requiring NVIDIA GPUs of Ampere generation or newer. Released checkpoints are organized by scale and pretraining budget — base_45m_200h, base_45m_1kh, big_350m_200h, and big_350m_2kh — with per-task fine-tuned variants for human (H1) and monkey (M1, M2) evaluation splits, along with the normalization statistics needed to reproduce them. Weights are distributed through Hugging Face under a CC BY-NC 4.0 licence, while the training and evaluation code on GitHub is MIT-licensed. The training stack uses PyTorch Lightning with Weights & Biases logging, hyperparameter sweeps, and SLURM submission for multi-node runs. Evaluation reports kinematic R² against held-out behaviour, the standard metric for continuous motor decoding.

#Applications

The immediate beneficiary is clinical BCI research, where human recording time is the binding constraint: pretraining on pooled monkey and human data reduces how much per-participant data is needed to reach a usable decoder. Beyond assistive neurotechnology, the released checkpoints give systems neuroscientists a general-purpose representation of motor cortical population activity that can be probed, fine-tuned, or used as an initialization for new experimental paradigms. Because the pretraining corpus spans many labs and array types, the model is also a useful testbed for studying what does and does not transfer across intracortical recordings.

#Impact

NDT3 was published at NeurIPS 2025 and is among the largest empirical tests of scaling laws for intracortical neural data. Its most useful contribution may be negative: by pretraining across 2,000 hours and 350M parameters and still finding sensor variability and stereotyped behaviour to be limiting, it redirects attention from model scale toward data diversity and recording standardization. The non-commercial licence on the weights restricts industrial reuse, and the Ampere-or-newer hardware requirement raises the floor for running the larger checkpoint.