FastVLM
Apple's efficient vision-language model for real-time on-device multimodal AI.
About
FastVLM, from Apple and published at CVPR 2025, is built around FastViTHD, a hybrid vision encoder designed to emit far fewer visual tokens and slash encoding latency on high-resolution images. Apple reports the 0.5B variant beats LLaVA-OneVision-0.5B with 85x faster time-to-first-token and a 3.4x smaller encoder, while the 7B variant competes with Cambrian-1-8B at nearly 8x faster token generation — aimed at real-time inference on iPhone-class hardware.
Code, checkpoints, and an iOS demo app are open-sourced via apple/ml-fastvlm on GitHub with weights on Hugging Face. The exact license on the repo should be confirmed before commercial use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported (HF model card): FastVLM-7B scores 83.6 Ai2D, 93.2 DocVQA, 45.4 MMMU; 0.5B variant at 85x faster time-to-first-token vs LLaVA-OneVision-0.5B.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: CVPR 2025; open code, checkpoints, iOS demo.; Efficiency numbers are Apple-reported; third-party replication limited.. License: Open-sourced by Apple; confirm exact repo license before commercial use..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · CVPR 2025; open code, checkpoints, iOS demo.
- · Efficiency numbers are Apple-reported; third-party replication limited.