Show HN: Peekaboolean – image and jev-like typed questions in typed anwers out
Category: ai-ml
Tags: vision-language-model, vqa, calibration, lora, structured-output, local-inference
Score: 8.0/10 (Innovation: 8, Technical: 8, Documentation: 9, Utility: 7)
Peekaboolean is a small vision-language model that answers typed questions about images (choice, score, yes/no) by scoring caller-supplied options via logit differences rather than free-form text generation. It supports calibrated probabilities, abstention, multiple serving modes, and achieves laptop-friendly latency (248ms p50 on M1 Max). The project also includes a larger reasoning variant and is built on LiquidAI's LFM2.5-VL-450M with LoRA fine-tuning.
Target audience: ML engineers, backend developers, data engineers, researchers
Repository: https://github.com/bykof/peekaboolean · Python · Apache-2.0 · 4 stars
View on Hacker News