Laya
Small encoder models that answer routing and triage questions in one pass
Emerged finds new products worth a look, daily — what’s new today
If you are paying a chat model to answer 'is this spam?' and then parsing the sentence back into a boolean, SemIf does it in one forward pass on your own GPU.
SemIf, formerly OpenJev, is Theodore Lee's open reimplementation of the Jev pattern: hand a model a question and a fixed set of options, and instead of sampling a reply it reads the probability of each option from the logits. One pass, calibrated numbers, no JSON to repair. It runs on CUDA, on Apple Silicon via MPS, and on CPU through llama.cpp; a browser demo at openjev.com loads a 0.6B to 4B model on-device so you can compare logit reading with token-by-token JSON. MIT, 3.5k stars in weeks, a 27B EXL3 bridge added this month. It copies the interface, not TypeSafe's model, and says so.
More like this
Bespoke Nimble
An open 9B model, dataset and training recipe for fast typed decisions over text
Kev
Open 0.8B to 9B decision models for choice, yes/no and rating questions
NanoJev
A 0.6B decision model that plays ViZDoom, mazes and Snake without decoding
jevals
Run dozens of agent evals per trace in one sub-second request, no judge prompt
jev-align
Teach a Jev classifier your labels by reviewing only its least certain answers