Breaking

Apple Releases MobileCLIP2 for Faster On-Device Image-Text Models

September 16, 2026 at 09:46 EDT

Foundation Models
Open Source

Apple has released MobileCLIP2, a new family of image-text models and training code aimed at low-latency multimodal understanding on devices. The headline result is MobileCLIP2-S4: Apple says it matches SigLIP-SO400M/14 on ImageNet-1k zero-shot accuracy while using 2x fewer parameters and beating DFN ViT-L/14 at 2.5x lower latency on an iPhone 12 Pro Max.

Sep 2025 · Apple

MobileCLIP2 moves CLIP to the edge

Apple's new image-text model family targets low-latency on-device understanding, with S4 matching SigLIP-SO400M/14 accuracy while using fewer parameters.

81.9%

S4 ImageNet-1k zero-shot top-1

Fewer parameters than SigLIP-SO400M/14

2.5x

Lower latency than DFN ViT-L/14

Released MobileCLIP2 variants

Latency drawn to scale

Reported image plus text latency on iPhone 12 Pro Max: 26.2 ms for MobileCLIP2-S4 versus 64.5 ms for DFN ViT-L/14.

26.2 ms

MobileCLIP2-S4

64.5 ms

DFN ViT-L/14

The shorter S4 column is the practical story: similar ImageNet accuracy with much lower reported latency.

How the recipe works

DFN data

→

Teacher embeddings

→

Synthetic captions

→

MobileCLIP2

Why developers care

Same-size upgrade paths, quantization targets and local image search or tagging workflows are natural fits.

What still matters

Retrieval performance is not always state of the art, and code, model weights and data use separate terms.

Continue reading

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