Fartash Faghri

Career

Faghri completed a PhD in computer science at the University of Toronto, supervised by David Fleet, with doctoral research on fast and robust training methods. During the PhD, Faghri interned at Google Brain in 2017 and 2020. Faghri joined Apple in 2021 and works in Apple Machine Learning Research.¹⁵

Research

Faghri's early work includes VSE++ (2017), on visual-semantic embeddings trained with hard negatives. At Apple, Faghri has worked on efficient image-text models, including MobileCLIP (2023) and its successor MobileCLIP2, whose models, training code and data-generation code were released in 2025, and on continual training of CLIP models in TiC-CLIP (2023).⁶⁴⁷¹

Faghri is the second author of FastVLM (2024), presented at CVPR 2025, which targets lower latency in vision-language models at high image resolutions. Faghri also co-authored DataComp-LM, presented at NeurIPS 2024, and co-organized a NeurIPS 2025 workshop on continual and compatible foundation model updates.³⁸¹

Work and contributions

• FastVLM Open weights

  Second author of the FastVLM paper, which introduces a hybrid vision encoder that reduces encoding latency and the number of visual tokens for high-resolution inputs to vision-language models.³⁹

• FastVLM 7B Open weights

  Listed among the authors in the citation of the ml-fastvlm repository, which provides the code for the FastVLM models.⁹

• DataComp-LM (DCLM) Open data

  Listed among the authors of the DataComp-LM paper on a testbed for controlled dataset experiments for language models.⁸¹

Timeline

1. Jul 2017Posted VSE++, on improving visual-semantic embeddings with hard negatives, as first author.⁶
2. 2021Joined Apple after completing a PhD at the University of Toronto.¹
3. Oct 2023Co-authored TiC-CLIP, on continual training of CLIP models.⁷
4. Nov 2023Co-authored MobileCLIP, on fast image-text models through multi-modal reinforced training.⁴
5. Jun 2024Co-authored the DataComp-LM paper on dataset experiments for language models.⁸
6. Dec 2024Posted FastVLM, on efficient vision encoding for vision-language models, as second author.³
7. Aug 2025Released MobileCLIP2 models with training and data-generation code.¹

Where they work

• Apple 27 people in the wiki

Related people

• Vaishaal ShankarAlso at Apple
• Vipul Ved PrakashAlso at Apple
• David GrangierAlso at Apple
• Alexander ToshevAlso at Apple
• Ruoming PangAlso at Apple
• Fei-Fei LiDatasets and data · Multimodal models
• Matt DeitkeMultimodal models · Datasets and data
• Ce ZhangDatasets and data · Training systems
• Sara HookerDatasets and data · Training systems
• Peter StaarDatasets and data · Multimodal models