400M model handles game NPC judgment calls in ~0.1s on a 3060 Ti, Apache-2.0
NoahPersaud · reddit · 2026-11-10
A developer released gamedecider (Apache-2.0), an Ettin-400M encoder fine-tuned for quick game NPC judgment calls: given a scene, question and options, it outputs probabilities or "not sure". Designed to sit beside a local LLM that writes dialogue while the small model cheaply handles yes/no and pick-one checks (threat detection, apology sincerity, secret leakage), or acts as a checker over candidate replies. 0.1s per check on a 3060 Ti, 1s on CPU; up to 5 questions per scene cost the same as one. Ships 42 adapters (threat, apology, persuasion, Skyrim-style memory matching, NPC mood), pip-installable, fully local. Weaknesses: slow CPU on long scenes, English-only, planning untested beyond own generated data, no public benchmarks yet. https://huggingface.co/holdoutlabs/gamedecider
More from Models
- Investigation: frontier coding LLMs reward-hack by recovering answers, even in hardened settings — logitly-ai · 2026-11-10
- Australian lab Maincode releases Qwen3-stitched Matilda Platypus base models — ThePrimeClock · 2026-11-10
- Cortex hits 73.2% on τ³-bench banking, 18 points above best published model — multiply_matrix · 2026-11-10
- Anthropic says it deliberately capped Model 3's AI-accelerating research capabilities — ziv_ravid · 2026-11-10
- Blogger benchmarks three top models: Opus for code, Fable for architecture — aigclink · 2026-11-10
- Plus Subscriber Complains: No Free Resets, Out of Usage for Three Days — TameYour · 2026-11-10