AIglance-vlm speedlab goes open source for local webcam AI detection
The announcement describes emotion, counting and object detectors that run locally, and reports 27.6% lower median latency in a controlled nine-question benchmark.
TLDR
The project’s announcement describes glance-vlm speedlab as a way to run multiple live webcam AI detectors locally. In a live one-question loop on an Apple M5, it reports median (p50) latency of roughly 210 ms with PyTorch/MPS FP16 and 160 ms with MLX 8-bit. In a separate controlled nine-question benchmark using fresh frames, it reports p50 latency falling from 358.5 to 259.6 ms—a 27.6% reduction—with 84 of 84 decisions matching. The announcement also lists 21 experiments, reproducible benchmarks, a paper and failures.
