Gensyn releases open-1b, the first fully auditable open-source model

On September 15, decentralized-compute company Gensyn released open-1b, a 1B (≈1.6B) parameter open-source language model shipped with "proof of training": fingerprints for every training step plus tooling so anyone can bitwise replay and audit the run on their own hardware. The team and sharers position it as the first end-to-end auditable model built on "don't trust, verify".

Confirmed

Spec: 24-layer decoder-only transformer, roughly 1B/1.6B parameters
Every training step is bitwise reproducible; training data is fully open and traceable to the exact step each data point entered
Any checkpoint can be re-run with custom kernels and verified on both NVIDIA and Apple hardware
Provides proofs for data, recipe, biases and weights, claiming coverage of every step of training and inference — the first fully auditable transformer training run
Core design: training itself stays centralized (decentralizing it is too costly), but verification is decentralized

Why it matters

@benfielding argues debates over regulation, slowing down, and open weights all reduce to verifiability; open-1b shows AI systems can be verified without trusting the publisher or repeating the entire training run
He also notes, referencing a Time cover, that public distrust of AI is addressable: full traceability gives an independently checkable tool for questions about data provenance, recipe and bias, rather than relying on the publisher's word
Authors see verifiable training as a key tool against future AI disputes, contrasting with the closed approach of mainstream AI labs

2026-09-15 ~ 2026-09-17 · 9 related posts

Primary sources

OPEN-1B: the world's first fully auditable 1.6B-parameter transformer training run — benfielding · 2026-09-15 23:18 (UTC+08:00)
Gensyn ships open-1b, the first language model with bit-for-bit auditable training proof — benfielding · 2026-09-15 23:20 (UTC+08:00)

[source] OPEN-1B: the world's first fully auditable 1.6B-parameter transformer training run — benfielding · 2026-09-15
[source] Gensyn ships open-1b, the first language model with bit-for-bit auditable training proof — benfielding · 2026-09-15
GenSyn Releases OPEN-1B: A Fully Auditable 1B-Parameter LLM With Bit-for-Bit Replayable Training — benfielding · 2026-09-15
Gensyn's open-1b: first model with bitwise-reproducible, fully auditable training — benfielding · 2026-09-16
Gensyn releases open-1b, the first fully auditable and replayable foundation model — flngr · 2026-09-17
Gensyn launches open-1b, billed as the first model with fully auditable training — benfielding · 2026-09-17

3 near-duplicate retellings: benfielding · micoolcho · benfielding