Open-1b publishes checkpoints, data, code, and hashes for every training step. Gensyn says audits work across several kinds of hardware, while its reproducible runtime is about five times slower than an optimised system.
New to this? Read it in simple words
- Gensyn released a small AI model called open-1b, with its code, data and checkpoints. People can replay any of its 80,957 training steps and compare the results.
- Gensyn says this method is about five times slower than an optimised PyTorch system. So this clear evidence comes with a real speed and energy cost.
- Open weights alone show only the finished model. This release can also show how the model was made.
- A matching result does not prove the model is safe, fair or useful. Independent teams still need to test the full evidence.
- Open weights
- The trained model files are public, so anyone can download and run the model.
- Checkpoint
- A saved copy of an AI model at one point during its training.
- Hash
- A short code made from data. If the data changes, the code also changes.
The release tries to connect model weights to their full history
Open model weights show what a model became. They do not prove which data and operations created those weights. Gensyn adds a public hash for the data, model state, optimiser state, and gradients at every training step.
An auditor can choose a step, load the earlier checkpoint, and replay the work. The new result should match the published hash. Gensyn says this works on Nvidia hardware, x86 and Arm processors, and Apple Silicon.
The release includes base and instruction models, training code, data, checkpoints, an audit tool, and a dataset search tool. This gives outside researchers more evidence than a normal open-weight release.
Open-1b adds evidence beyond open weights. Gensyn reports that its reproducible runtime is about five times slower.
Reproducibility required strict and slower computing rules
Normal processors can add the same numbers in different orders. Tiny differences then grow during training. Gensyn fixed the order of operations, random choices, data delivery, and communication between machines.
The model has 1.61 billion parameters and used hundreds of billions of training tokens. Gensyn reports that the main run took 27.8 active days on 48 Nvidia H100 processors. The full work lasted about 29.5 calendar days.
The reproducible system reached about five percent model computing use. Gensyn says this is about five times slower than an optimised PyTorch system with the same recipe. Clear evidence therefore comes with a real speed and energy cost.
Sources1
A public audit needs independent people to use it
Gensyn invites people to replay any of 80,957 steps. One laptop cannot check the whole run quickly, so the project records checks from many machines. The public site shows which parts have been checked.
Matching a hash can confirm that a published step follows the published recipe. It cannot prove that the model is safe, fair, or useful. Reviewers must still study the data, licence, test results, and possible harmful behaviour.
The release is an important technical experiment, not a final standard. Independent teams should reproduce random steps, examine missing evidence, and publish failures. Larger models will also show whether the method can scale at a practical cost.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · September 15, 2026Introducing open-1b: the first model you do not have to trust Gensyn
- 2
- 3Primary source · September 15, 2026Gensyn challenges Big Tech with an auditable AI model PR Newswire
We read Gensyn’s technical release and opened its public audit site. We used the timed press release to confirm the launch. The design, speed, and scale figures are Gensyn claims until independent auditors publish their own results.