This 106M-param model plays Doom at 10ms on a single CPU core, beating Jev

Playful_Suggestion_3 · reddit · 2026-09-29

The author released indecis, an MIT-licensed Go library that fine-tunes a small text encoder (106M params, tens of MB) to answer typed questions with probabilities, letting it play Doom at 10ms median latency on a single CPU core — versus Jev's 70–500ms hosted decisions and LAYA's 421M params at 98ms. Across 20 ViZDoom defendthecenter rounds it scored +21.8 points on average, matching the scripted bot that generated its training data, using 65MB of RAM and just 11 minutes of laptop training. The real goal: replacing the LLM-based request classifier in the author's Xolo gateway to cut hundreds of ms of network round-trip.

Original post →

More from Models

Models channel →