Bench Labs released the compact Cagliostro-v3 model

Company Bench Labs introduced Cagliostro-v3 — a new decoder-only model optimized for work with limited computational resources. The system architecture has 146 million parameters, which refers it to the category of compact language models (SLM).

The pre-training process of the model was carried out on a data array of 75 billion tokens. At the moment, training has been completed on 72.7 billion tokens. The developers assigned the model an Open SLM Index of 26.27, which allows evaluating its effectiveness in comparison with similar solutions on the market.

During testing, Cagliostro-v3 demonstrated leadership in the ArithMark-3 benchmark, specializing in checking the mathematical abilities of neural networks. At the same time, in tests for logical thinking and common sense, such as PIQA, the model shows results below the expected indicators for its class.

The release of the model is aimed at developing the segment of lightweight systems capable of performing specialized tasks without the need to deploy heavy infrastructure solutions. The source code and model weights are already available for download in the repository on the Hugging Face platform.