Show HN: VisionLaya: Jev with Vision capabilities

Category: ai-ml

Tags: vision-language, calibration, efficient-inference, game-playing, smolvlm

Score: 6.3/10 (Innovation: 7, Technical: 7, Documentation: 7, Utility: 5)

VisionLaya is a 201M parameter vision-language model that makes typed, calibrated decisions (multiple choice, yes/no probability, rubric score) from an image and optional text in a single forward pass without text generation. It is notable for its efficiency (2 hours training on one H100), calibration, and ability to play simple games from pixels, found via an autoresearch loop.

Target audience: machine learning engineers, AI researchers, developers building vision-based decision systems

Repository: https://huggingface.co/thaitea/laya-vision-smolvlm-256m

View on Hacker News