Gensyn released a compact model open-1b-sft

Company Gensyn introduced a new language model open-1b-sft designed for chat tasks. The model has 1.61 billion parameters and underwent a procedure of further training (SFT) on the Tulu-3 dataset.

A key feature of the release is the complete verifiability of the training process. The developers provided the ability to audit all stages of model preparation, which allows tracking the correctness of weight formation. All code and data for reproducing the results are available in open access on the Hugging Face platform.

The technical characteristics of open-1b-sft are optimized for operation on limited computing resources. The model demonstrates an evaluation of 9 points on the internal quality metric, achieving this result in 1 hour of training on specialized infrastructure. The architecture is focused on high-speed response generation while maintaining basic dialogue skills.

The release of this model expands the ecosystem of compact solutions that can be deployed locally without significant resource costs. The transparency of the training process makes open-1b-sft a tool for research in the field of result reproducibility and neural network verification.