THX-01: open 322M decision model classifies tickets in ~10 ms

The team open-sourced THX-01, a 322M-parameter non-autoregressive model (mmBERT-base encoder plus a 15M decision head) under Apache 2.0. It answers typed questions (choice, yes/no, score, number, excerpt) in a single forward pass at roughly 10 ms per request on one GPU, or about 3,000 decisions/s batched, and also runs on CPU.

• 322M parameters: mmBERT-base encoder plus 15M decision head, Apache 2.0
• 98.4% accuracy on a 2,843-ticket benchmark vs 98.5% for Claude Sonnet 5.5
• ~10 ms latency on GPU vs 1.5 s for Sonnet 5.5, also runs on CPU
• Calibrated probabilities (ECE 0.003): ~75% of tickets handled fully automatically with zero errors