Show HN: THX-01 – Open- source decision model that runs on phones and beats Jev
Category: ai-ml
Tags: decision-making, calibration, multilingual, non-autoregressive, reinforcement-learning, information-extraction
Score: 7.3/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 7)
THX-01 is a 322M parameter non-autoregressive decision model that answers typed questions (choice, number, excerpt, citation) over documents in a single forward pass with calibrated probabilities. It uses reinforcement learning with a strictly proper scoring rule to ensure honest confidence, enabling selective automation. Benchmarks show it matches or beats larger LLMs on classification and extraction tasks while running in ~10ms on one GPU.
Target audience: backend developers, data engineers, AI/ML engineers
Repository: https://huggingface.co/doofz/THX-01
View on Hacker News