NanoJev
A 0.6B decision model that plays ViZDoom, mazes and Snake without decoding
Emerged finds new products worth a look, daily — what’s new today
Bespoke Nimble
An open 9B model, dataset and training recipe for fast typed decisions over text
It is the most complete public recipe yet for building your own Jev-style classifier: the data method, the held-out scores and the serving code are all there.
Nimble takes a context and a flat schema of enum or true/false questions, reads the prompt once and scores a single answer token per question, returning the choice and each option's probability with no JSON to parse. Bespoke Labs fine-tuned Qwen3.5-9B with LoRA on data built by flipping one fact so the right answer changes. On 324 held-out examples it matches 90.1% of reference labels, against 66.4% for the base model and 93.2% for Jev. It runs on Apple Silicon or NVIDIA, with training data, a temperature fit and public benchmarks added since launch.
More like this
Kev
Open 0.8B to 9B decision models for choice, yes/no and rating questions
SemIf
Read a yes/no probability straight out of a small open model, no generated text
jev-align
Teach a Jev classifier your labels by reviewing only its least certain answers
jevals
Run dozens of agent evals per trace in one sub-second request, no judge prompt
Laya
Small encoder models that answer routing and triage questions in one pass