FocsleFo’c’sle
AP

MobileCLIP-B

5%
by Apple

Larger MobileCLIP. Better recall, twice the compute.

MultimodalApache-2.0FP16clipembedding
142K downloads 8.2K deploymentsUpdated Feb 4, 2028
Headline:9.8ms · Apple Neural Engine (M4) · FP16
OverviewBenchmarks5Sim ResultsDeploy5FilesDiscussion23

About this model

Larger MobileCLIP. Better recall, twice the compute.

Authored by apple. Curated into the Fo’c’sle reference set on 2028-02-04. All cross-chip benchmarks below were collected in matched-pair runs in the HIL lab using the same input pipeline, same upstream preprocessing, and the same downstream consumer. See the methodology page for the full protocol.

Task
Multimodal
Parameters
86.3 M
Benchmarked on
5 chips
Deployments
8.2K

Architecture

Vision-text dual encoder
Inferred from upstream weights · simplified
ImageTextVision towerText towerProjectionEmbedding

Headline benchmarks

NNVIDIA Jetson AGX OrinFP16
28.7ms p50
35 FPS99.5% acc42.2 W
QSnapdragon 8 Gen 3 NPUFP16
61.5ms p50
16 FPS99.6% acc7.4 W
Apple Neural Engine (M4)FP16
66.9ms p50
15 FPS99.9% acc6.6 W

Training data

Pretrained on the upstream maintainer’s released checkpoint. Edge-distillation pass uses 2.4M frames from the Fo’c’sle distillation corpus (consented public data + opt-in publisher contributions). Quantization-aware fine-tune uses 320K calibration samples drawn from the target task’s eval domain.

  • Pretraining corpus: upstream maintainer release
  • Distillation corpus: 2,400,000 frames
  • Calibration set: 320,000 samples (per task)
  • Eval set: standard benchmark + matched-pair HIL runs