Benchmarks & research
Nandakishor M's Vega answers choice, score and yes/no questions by letting a particle roll downhill into one valley per answer, on top of a frozen Qwen3.5 reader.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Its trained part is a 57 MB engine for the 0.8B version, or 124 MB for the 4B. Where and how the particle comes to rest sets each probability, and can raise an abstain or unbound flag instead of an answer. It is separate from Laya, his earlier model.
Vega suits sorting jobs where the possible answers are known in advance, such as checking emails for scams, when the text must stay on your own computer. It can also say it is unsure, or that a question is outside what it knows; the maker says it often does this with unfamiliar kinds of question. Treat those answers as a cue for a person to look.