Gensyn Releases open-1b, an Apache 2.0 Model With Auditable Training Hashes

Gensyn announced open-1b on September 15, 2026, releasing a 1.61B-parameter decoder-only language model under Apache 2.0 with a public verification record that commits a state hash for each of 80,957 optimizer steps. The Hugging Face card for Gensyn/open-1b-base states the model was pretrained on 400B tokens and ships with intermediate checkpoints, training code, dataset access, and an audit path designed for independent bit-level replay.

Unlike open-weight drops that publish only final parameters, open-1b chains each step’s hash to the previous one and publishes Merkle roots so auditors can check inclusion without trusting the record’s backend. Gensyn also published gensyn-audit on PyPI under Apache 2.0 — first as v0.1.0 on September 14, 2026, then 1.0.0 on September 15 and 1.1.0 on September 16 (PyPI latest at check time) — and points auditors to open1b.gensyn.ai for claimable step audits. The current audit client documentation says Apple Silicon Macs and Python 3.11 are required.

What shipped

The open-1b family on Hugging Face includes open-1b-base, open-1b-midtrained-93B, and open-1b-sft. Supporting repositories named on the card include github.com/gensyn-ai/open-transformers for training code and github.com/gensyn-ai/gensyn-audit / PyPI gensyn-audit for the replay CLI. Treat any capability claims beyond the documented pretraining setup as outside this brief; the release’s distinguishing claim is auditable training provenance, not a new leaderboard sweep.

Primary sources are the September 15 Gensyn PR Newswire announcement, the Gensyn/open-1b-base Hugging Face card, the PyPI gensyn-audit project page, and the public verification site at open1b.gensyn.ai.