THX-01: An Open-Source Decision Model That Runs on Your Phone and Beats Jev (doofz.substack.com)

HAL-X AI has announced THX-01, a groundbreaking open-source decision model designed to enhance efficiency in AI task handling. Unlike traditional large language models that generate text for simple tasks, THX-01 directly processes questions and documents in a single forward pass, achieving around 10 ms inference on GPU. With a compact architecture of 322 million parameters, it can run locally on devices such as smartphones, supporting 20 languages. This capability opens up new avenues for multilingual applications and on-device processing, making AI accessible without significant infrastructure. THX-01 outperforms its predecessor, Jev, in several decision-making tasks while introducing unique features like native number extraction, verbatim excerpts, and calibrated confidence estimates. Its performance is also impressive, achieving 98.4% accuracy in support ticket classification and can automate approximately 75% of tickets without error. By providing this fully open-source model under the Apache 2.0 license, HAL-X AI is paving the way for a shift in AI infrastructure towards smaller, specialized models that prioritize quick, reliable decision-making over extensive generative tasks. This release could spur innovation in applications where precise, efficient decisions are paramount.