Skip navigation
glance-vlm speedlab is now open source!
0:00 / 1:41

glance-vlm speedlab is now open source!

Yohei

@yoheinakajima126K followers

Visit on X
288 likesDownloadOriginal post

51,739 views

glance-vlm speedlab is now open source! read: try: turn any webcam into multiple live AI detectors (emotion, count, object), running locally ⚡ Recorded live on an Apple M5, one-question loop: PyTorch/MPS FP16: ~210 ms p50 MLX 8-bit: ~160 ms p50 🧪 Controlled nine-question, fresh-frame benchmark: 358.5 → 259.6 ms p50 27.6% lower latency 84/84 decisions matched 21 experiments, reproducible benchmarks, paper, and failures

27 Comments

Comments from the original post on X (Twitter)

  • Yohei's profile picture

    @yoheinakajima

    glance-vlm is a harness for measuring what happens when you chop off the generation from an open source vlm and use it as a classifier/decision model:

    4 likes

  • Yohei's profile picture

    @yoheinakajima

    ICYMI that’s 161 ms latency on general vision true/false Qs at the end, running locally on webcam

    8 likes

  • Yohei's profile picture

    @yoheinakajima

    compare glance 4B to laya-vision

    3 likes

  • Assaf Elovic's profile picture

    @assaf_elovic

    Nice. One use case I see is to monitor elderly. My grandma is quite sick and it takes a lot of time, energy and attention to make sure she is good. Would be awesome to get an alert once the camera notices something weird in case we are not there. What do you think? Can be done?

    1 like

  • Yohei's profile picture

    @yoheinakajima

    I think worth playing around with! Once you know a good use case you might want to go specialized but this is an easy way to prototype something quickly

  • John Rood's profile picture

    @johnroodepic

    for the temporal gate, compare against the last scored frame, not the previous camera frame. otherwise slow drift can stay below the threshold forever while the answer gets stale.

    1 like

  • Yohei's profile picture

    @yoheinakajima

    hm thanks will take a look at this

  • chetan conikee's profile picture

    @conikeec

    Really neat . love it

    1 like

  • Yohei's profile picture

    @yoheinakajima

    thx :)

  • ZAZA's profile picture

    @OpinionAILtd

    This is game changer for vision AI

    1 like

  • Tom Bielecki's profile picture

    @tombielecki

    🐐

    1 like

  • Shiva's profile picture

    @LShivkumarreddy

    This is good!

    1 like

  • Ran Aroussi's profile picture

    @aroussi

    This is awesome! I need to fork it and turn it to a meeting companion that helps neurodivergent people with reading emotions and social queues.

    1 like

  • Jon Kraayenbrink's profile picture

    @kraayenJon

    this is so cool!

    1 like

  • Yohei's profile picture

    @yoheinakajima

    Thx :)

    1 like

  • Emma W's profile picture

    @rayane_esteves

    210ms locally is the bit that jumps out, that puts weird webcam workflows in reach without renting a gpu every time. curious how hot the m 5 gets after 20 minutes though

    1 like

  • Ilia Pomirchi's profile picture

    @FantomKoenig

    Agents got hands in 2023. Eyes took until 2026, at 260ms a glance. Counting people in a frame is arithmetic, a face is a guess. Which of your 21 failures surprised you most?

  • Joe Devon's profile picture

    @joedevon

    Great work Yohei!

  • Mira Takes's profile picture

    @meimu7ns4

    The interesting bit isn’t the webcam demo—it’s the loop shape. Turning a VLM into a cheap local decision engine is a much more useful product primitive than another chat wrapper.

  • Mouten📡's profile picture

    @Mouten_dev

    glance-vlmのspeedlab OSS、僕は触りたい。 webcamをローカル検出に、って相手の切り口は受け取る。 自分では未開。まずは