Jev Cards
A picture-card quiz where you race a vision model that answers in 0.3 seconds
Emerged finds new products worth a look, daily — what’s new today
Jev Plays Pokémon Red
A cheap decision model took 16,150 turns and $1.65 to reach the Hall of Fame
It is a cheap, fully logged way to see what a small decision model can do when a good harness handles pathfinding and the model only chooses.
A complete, recorded run of Pokémon Red in which Jev, a fast decision model, picks every move from options a TypeScript harness lays out. The run took 37 hours 40 minutes, 16 team wipes and 15 Elite Four attempts, at roughly 0.4 seconds and a fraction of a cent per decision. Code is GPL-2.0 on GitHub, footage is on YouTube, and no ROM is included. Commenters point out that the harness does much of the work.
More like this
NanoJev
A 0.6B decision model that plays ViZDoom, mazes and Snake without decoding
jevman
An arcade maze turned into a latency test for decision models
jevals
Run dozens of agent evals per trace in one sub-second request, no judge prompt
Jev Ultrafast
An open browser agent that finishes a flight search in about seven seconds
Bespoke Nimble
An open 9B model, dataset and training recipe for fast typed decisions over text