The journal
Engineering / September 18, 2026 / 4 min read

Small should be a measurement.

A model’s download size is only one part of its footprint.

Count the complete cost.

The artifact sizes in our catalog are the raw UTF-8 JSON files containing labels, features, and weights. They exclude application code, runtime memory, and any compression applied by a server. We report exactly what we measured so the numbers stay useful.

Different tasks need different models.

A Gaussian classifier over five engineered audio features can occupy less than a kilobyte. That does not make it equivalent to a neural speech model. It reflects the narrower representation and the much simpler training problem.

Make releases reproducible.

The training script uses fixed seeds and deterministic serialization. A manifest records the SHA-256 digest of each artifact. The SDK reads an explicit artifact instead of silently fetching new weights during inference.

Measure the application too.

Before releasing a production SDK, we need device-level latency, peak memory, cold-start behavior, and task quality on independent datasets. Our research preview does not yet claim those measurements.

Explore the research models