Breaking

Apple Publishes MobileCLIP2, Reporting SigLIP-Class Accuracy at Half the Parameters

September 23, 2026 at 10:54 EDT

Foundation Models
Open Source

Apple has published MobileCLIP2, a family of efficient image-text models designed to move CLIP-style visual understanding closer to phones and other edge devices. The company reports that MobileCLIP2-S4 matches SigLIP-SO400M/14 on ImageNet-1k zero-shot accuracy while using about half the total parameters, and runs at 2.5x lower latency than DFN ViT-L/14 in measurements on an iPhone 12 Pro Max.

Aug. 28, 2025 | Apple

MobileCLIP2 shrinks image-text AI

Apple reports SigLIP-class ImageNet accuracy with a smaller, edge-oriented model family built for fast image-text understanding.

81.9%MobileCLIP2-S4 ImageNet-1k top-1

2xFewer parameters than SigLIP-SO400M/14

2.5xLower latency than DFN ViT-L/14

6MobileCLIP2 checkpoints listed

Parameter scale at matched accuracy

MobileCLIP2-S4 reaches 81.9% ImageNet top-1; SigLIP-SO400M/14 is listed at 82.0%.

MobileCLIP2-S4445.2M params

SigLIP-SO400M/14877.4M params

Each segment is about 100M parameters.

How Apple changed the recipe

DFN data

Higher-quality base filtering

Teacher ensembles

Distillation from CLIP models

Synthetic captions

CoCa-generated text variants

Mobile checkpoints

S0 through S4 variants

Why developers care

Small image-text embeddings can power local photo search, screenshot search, zero-shot labels and fast VLM encoders.

Where it still pinches

The paper flags retrieval tradeoffs, and Core ML conversion can require fixes for dynamic shapes, attention and reparameterization.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.