NEST

Models

Checkpoints and releases

None released yet. Here is what will ship and how to know when.

No checkpoint exists yet. NEST v2 has not been trained to completion, so there is nothing to download. An earlier version of this page carried a model table with word error rate numbers. That table was removed: those numbers were estimated from training loss, never measured by decoding, and were never real results. A checkpoint appears here only after it is trained and evaluated with free decoding on held-out subjects.

Planned releases

What ships, once it exists

Both NEST v2 heads instantiate and run today. Neither has completed training. Parameters below are counted from the code, not estimated.

ModelParametersInputStatusWER
NEST v2 CTC45.3M840-dim word featuresPlanned after GPU runpending
NEST v2 BART184M total, 64.7M trainable840-dim word featuresPlannedpending

When a checkpoint is ready, it will publish on GitHub Releases and the Hugging Face Hub, together with the evaluation report that produced its WER.

Get notified

Two ways to hear about it first

Watch the repository

Watch releases on GitHub. The v2 checkpoint and its evaluation report publish there the moment training finishes.

Follow Discussions

Training progress, the GPU run, and open questions about the evaluation protocol are posted to GitHub Discussions as they happen.

Train your own today

Nothing is stopping you from running it yourself

The code, the ZuCo integration, and the training loop are public now. A quick test run on the CTC head verifies your setup in minutes; a full run needs a GPU. See the docs for installation and the rest of the commands.

Quick test · CTC
python scripts/train_nest_v2.py --quick-test --model ctc

Follow the first checkpoint

Watch the repository for the v2 training run and the release that follows it.