Tilelli LLM: An Open-Source Anti-Hallucination Model Using Per-Token Routing

themoroccanship · reddit · 2026-08-06

Tilelli Lab has released Tilelli LLM, their first open-source model aimed at tackling LLM hallucinations. Instead of a monolithic architecture, it uses 3 specialized pathways for per-token routing. The team reports it catches out-of-distribution gibberish with a 0.93 AUROC and cleanly refuses unknown queries. A ternary version is also in active development to bridge the performance gap with standard float models. The project is fully open-source on GitHub and HuggingFace.

Original post →

More from Models

Models channel →