Kev
Open 0.8B to 9B decision models for choice, yes/no and rating questions
Emerged finds new products worth a look, daily — what’s new today
Researchers curious whether a tiny model can make fast typed decisions get the full pipeline, data and checkpoint to test it themselves.
NanoJev adds decision heads to Qwen3-0.6B so it returns probability distributions over supplied candidates, yes/no propositions or ordered levels without generating any tokens. Many states and questions are batched in a single forward pass. One checkpoint is trained across four game tasks and the README reports it beating TypeSafe's Jev on ViZDoom Basic (128/128 against 56/128) and on a 50×50 maze. Weights and an 18,760-question dataset are on Hugging Face; the interactive site is gated, but replays run locally.
More like this
Bespoke Nimble
An open 9B model, dataset and training recipe for fast typed decisions over text
SemIf
Read a yes/no probability straight out of a small open model, no generated text
Jev Plays Pokémon Red
A cheap decision model took 16,150 turns and $1.65 to reach the Hall of Fame
Jev Cards
A picture-card quiz where you race a vision model that answers in 0.3 seconds
Flide
Exact, shared-context inference server for small Jev-style decision models