Local gesture recognition with Liquid AI's d1-2B: ~70ms per frame, zero tokens generated
helloiamleonie · x · 2026-11-18
Developer ivanfioravanti open-sourced mlx-d1-gestures, which turns a webcam feed into live hand-gesture and facial-expression recognition on Apple Silicon using Liquid AI's d1-2B decision model via mlx-vlm.
- Zero tokens generated: each frame asks four typed questions (gesture? expression? hand location? face location?) answered in one forward pass — pure prefill, 70ms and 600 input tokens per frame on M5 Max, roughly 14 decisions/second, fully local with no frames uploaded
- New iOS/iPad apps: iPhone 18 Pro runs 180ms on ANE, 270ms on GPU; iPad Pro M5 230ms
- The author also fixed a bug where precompiled question caches forced a model redownload
It's a neat demonstration of repurposing a chat VLM as a low-latency local classifier; the updated repo is reproducible today.
More from Infra
- Meta scales up MTIA AI chip team, hiring across compilers, kernels and runtime — asdf1234_0 · 2026-11-18
- Smuggled Blackwell fingerprints found in Chinese open-weight models; Jensen Huang faces scrutiny — Difficult-Limit-7551 · 2026-11-18
- First SGLang Summit draws 2,000+: v1.0 ships, consortium forms with NVIDIA, Google, Meta — aliscodes · 2026-11-18
- New ApexW optimizer claims lower loss than AdamW with half the VRAM of bitsandbytes — omendUUde · 2026-11-18
- TokScale leaderboard: #1 user burned 4.65T tokens, est. $2.43M in cost — olcan · 2026-11-18
- MOSS-Image FP16-fixed weights bring speed boost to aging Turing GPUs — mkrGG · 2026-11-18