Qev-0.8B · New SOTA tiny local decision model (app.labqoat.com)

The recent release of the qev-0.8b tiny local decision model marks a significant advancement in AI, achieving state-of-the-art (SOTA) performance among models of similar size, with 77.1% accuracy across various benchmarks. Notably, it outperforms 26 other decision models with over 1 billion parameters, demonstrating not just the highest accuracy but also the fastest decision-making speed at 3.5 milliseconds per query. This model is particularly notable for its capacity to tackle seven different business tasks, including two focused on image interpretation, and five games—four of which were never included in its training data, showcasing its capability to generalize effectively. Designed to run efficiently on local devices, including Apple Silicon with only 1.3 GB of RAM required, qev-0.8b processes data without generating text, ensuring data privacy. It delivers results in calibrated probabilities, allowing users to set thresholds for confidence in decisions, and can function seamlessly in environments where quick decision-making is crucial. The model is also versatile, shipping in multiple formats for compatibility with various systems, making it accessible for diverse applications while maintaining a powerful inference ability without incurring additional token costs.