THX-01, a 322M open decision model that runs on phones, debuts on Hugging Face

HAL-X AI, an AI research company based in Azerbaijan, introduced THX-01 on the Hugging Face blog on October 10: a fully open-source decision model with 322 million parameters that can run locally on resource-constrained devices, including smartphones with a suitable runtime.

According to the post, THX-01 is non-autoregressive: it takes a state such as a document or message plus typed questions and returns structured answers with calibrated probabilities in a single forward pass, in roughly 10 milliseconds in the authors' measurements. It supports choice, yes/no probability, scoring, number extraction, excerpt extraction and citation.

On an internal multilingual support-ticket benchmark of 2,843 tickets, the authors reported 98.4% classification accuracy and about 33 times lower measured latency than TypeSafe Jev 1.13, though they noted the comparison used different serving conditions. The model weights and Python package are under Apache 2.0; a free hosted API with usage limits is also available.

The authors said THX-01 is designed for decisions, not open-ended text generation, and that numeric extraction operates only on values present in the source document.