Skip to content
← Models
Open weightsGoogle (Francois Belletti, Yu-Hui Chen, Ard Oerlemans, Ronny Votel)/10 trustMetrics-checked

MoveNet

Google's edge pose model: single and multi-person keypoints in the browser and on phones at real-time speed.

Visit official page →

About

MoveNet is Google's family of ultra-light pose models built for the browser (TensorFlow.js), Android (TF Lite), and TF Hub. A MobileNetV2 feature extractor feeds a feature-pyramid decoder and CenterNet-style prediction heads, outputting the 17 COCO keypoints with confidence scores in a single pass.

Lightning (192x192 input) is the speed pick, clearing 50 FPS on ordinary laptops; Thunder (256x256) trades speed for accuracy while staying real-time; MultiPose Lightning tracks up to 6 people at once. Google published full model cards under Apache-2.0, which is why it became the default pose backend for countless fitness, yoga, and dance-counting web demos.

Specs

Context window

Image/video input (192x192 Lightning, 256x256 Thunder, dynamic for MultiPose)

Pricing

Free — Apache-2.0 weights on TF Hub

Free quota

N/A (self-hosted weights)

API access

No hosted API — TF.js, TF Lite, and TF Hub weights for self-hosting

Weights

Open weights

Released

Not disclosed

Benchmarks

Provider-reported (official model cards): Lightning runs >50 FPS on most modern laptops, Thunder sustains real-time >30 FPS. Independent third-party paper (MovePose, arXiv 2308.09084) measured 53.6 AP (Lightning) and 64.8 AP (Thunder) on COCO val2017 — independent figures, not from Google.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

browser-pose-estimationmobile-pose-estimationfitness-trackingreal-time-keypoint-detectionedge-inference

Trust score — 8/10

/10 trustMetrics-checked

Scored 8/10 (metrics, metrics-v1). Evidence: official Google model cards with architecture details and licensing; Apache-2.0 weights on TF Hub; third-party independent AP figures consistent with Google's speed claims. Strengths: genuinely open weights and license, published specs, huge real-world use. Docked for: (1) no peer-reviewed paper, only model cards and a blog post; (2) accuracy benchmarks come from third parties, not Google. License: Apache-2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
6.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · Google (Belletti, Chen, Oerlemans, Votel), 2021; Lightning/Thunder/MultiPose variants for TF.js, TF Lite, TF Hub.
  • · Model-card claims: Lightning >50 FPS on laptops, Thunder >30 FPS real-time. Independent paper reports 53.6/64.8 AP on COCO val.
  • · No peer-reviewed paper — model cards + TF blog only; MediaPipe Pose is Google's newer focus.

Sources