AIMv2
Apple's multimodal autoregressive vision encoders, 300M to 3B.
About
AIMv2 is Apple's family of open-set vision encoders, published November 21, 2024 and added to Hugging Face Transformers in July 2025. It extends autoregressive vision pretraining into the multimodal setting: a vision encoder is paired with a decoder that generates raw image patches and text tokens, rather than relying on contrastive learning alone. Nineteen checkpoints span 300M to 3B parameters at 224, 336, and 448 resolutions.
The headline result is AIMv2-3B reaching 89.5 percent ImageNet-1k accuracy with a frozen trunk, and the authors report it consistently beating contrastive baselines like CLIP and SigLIP on multimodal understanding. Like other Apple research releases, the weights carry Apple's Machine Learning Research license, which is research-oriented rather than permissive open source.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
AIMv2-3B achieves 89.5% ImageNet-1k accuracy with frozen trunk; consistently outperforms CLIP and SigLIP on multimodal image understanding across settings (paper claims via HF Transformers model card)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics, metrics-v1). Evidence: Apple paper, HF apple/aimv2 collection with 19 checkpoints, Transformers integration July 2025. Strengths: strong published results from a major lab, wide size/resolution range, easy HF access. Docked for: (1) Apple Machine Learning Research license is research-oriented, not commercial-safe; (2) no official API or product support. License: Apple Machine Learning Research Model License (research use).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Apple, published 2024-11-21, Transformers integration 2025-07-08
- · 19 checkpoints, 300M to 3B; AIMv2-3B 89.5% ImageNet-1k frozen