AI Research Scientist (Multimodal Foundation Models)

Quick Facts

• Role: AI Research Scientist (Multimodal Foundation Models)

• Focus: Architecture, pre-training, and distillation for multimodal models

Description

You will investigate how multimodal foundation models should be designed, trained, and distilled—developing and evaluating new architectures, pre-training objectives, data strategies, optimization methods, and teacher–student learning techniques. A major focus is co-developing frontier models and efficient models for Apple silicon and on-device intelligence, incorporating deployment constraints (memory, latency, energy, privacy) throughout the research process. The work may lead to publications and open-sourced research artifacts, evaluations, or tools.

Responsibilities

• Research and develop multimodal foundation model architectures

• Design and test pre-training objectives, data mixtures, curricula, tokenization, and scaling/optimization methods

• Develop distillation methods that transfer capabilities from large to small models without treating distillation as an afterthought

• Co-train teacher–student systems and design offline/on-policy distillation using multiple supervision signals

• Evaluate models across reasoning, multimodal understanding, and instruction following

Requirements

• Hands-on experience running large-scale LLM pre-training experiments

• Experience with LLM pre-training topics: model architecture, training objectives, data mixtures, tokenization, curricula, scaling, and optimization

• Proficiency with PyTorch or JAX and distributed training systems

• Experience evaluating language understanding, reasoning, instruction following, and/or multimodal capabilities

• Strong understanding of transformer-based architectures and efficient/scalable foundation-model training

• Master’s degree (or equivalent practical experience) in machine learning, computer science, or related field

• Minimum 6 years relevant experience

Benefits

• Opportunity to shape new research directions with large-scale experiments

• Potential for product impact and a presence in the broader research community

• Possibility of publications and open-source release of selected models and research artifacts