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
L
Laya Vision
A 256M model that answers typed questions about an image with calibrated odds
When you need a yes-or-no, a choice or a 1-to-10 score about an image with a trustworthy probability attached, this does it in one forward pass at about 71 ms with no text generation to parse.
Laya Vision is an independent research fork of Laya, the calibrated decision model, that swaps its ModernBERT encoder for SmolVLM-256M so it can take an image alongside text. You pose typed questions, a choice, a rubric score or a yes-or-no, and get probabilities per answer rather than prose, in one forward pass at about 71 ms on an L4. Laya's predict API and calibration carry over unchanged. Across 26 validation sets it reports 74.1 percent accuracy with a calibration error of 0.034. Browser demo on Hugging Face. Apache-2.0 code, non-commercial weights. Opened 19 September, 23 stars.
More like this
W
wattai
A small local model that picks from your options and says how sure it is
Jev Cards
A picture-card quiz where you race a vision model that answers in 0.3 seconds
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
BlindSight
Webcam person detection where the cloud server never sees a single frame