AI Research Scientist: Multimodal Foundation Models - Architecture, Pre-Training & Distillation

Socket.dev; Seattle (WA); On-site; USD 150,000 - 210,000; Full time

Job summary

Apple seeks a Research Scientist to advance multimodal foundation models for on-device experiences. You will explore architectures, pre-training objectives, and distillation methods, aiming to balance capability with memory, latency, energy, and privacy requirements for Apple silicon.

In this role you will design and evaluate innovative model designs at scale, publish findings when appropriate, and translate successful ideas into practical technologies across Apple products and experiences.

Qualifications

• Hands-on experience designing, implementing, and running large-scale pre-training experiments for large language models.
• Strong proficiency with modern deep learning frameworks (PyTorch or JAX) and distributed training systems.
• Deep understanding of transformer-based architectures and scalable foundation-model training.
• Master's degree, or equivalent practical experience in machine learning or computer science, with 6+ years of relevant work.

Responsibilities

• Develop and evaluate new model architectures, pre-training objectives, and distillation techniques.
• Experiment at scale to translate ideas into foundation-model technologies for Apple products.
• Collaborate across model research, data, systems, and hardware to optimize on-device performance.

Skills

PyTorch

JAX

Distributed training

Transformer architectures

Machine learning

Education

Master's degree or equivalent practical experience in ML/CS

Job description

The Multimodal Intelligence Team is building the next generation of foundation models for Apple experiences. We are looking for a research scientist to advance the architectures, pre-training methods, and distillation techniques that make highly capable multimodal models practical across the Apple ecosystem. Our research spans the full foundation-model lifecycle: model architecture, pre-training objectives, data mixtures, optimization, scaling, distillation, and evaluation. A defining challenge of our work is to develop models that combine broad intelligence with the memory, latency, energy, and privacy requirements of on-device deployment. You will have the opportunity to shape new research directions, conduct ambitious experiments at scale, and translate successful ideas into foundation-model technologies that can reach Apple products. Where appropriate, this work may also lead to publications and the open sourcing of selected models, research artifacts, evaluations, or tools.

DESCRIPTION

In this role, you will investigate fundamental questions about how multimodal foundation models should be designed, trained, and distilled. You will develop and evaluate new model architectures, pre-training objectives, data strategies, optimization methods, and teacher-student learning techniques. Your work will explore how capabilities developed in large foundation models can be effectively transferred to smaller, more efficient models without treating distillation as an isolated downstream step. A major focus of the role will be the co-development of frontier models and efficient models for Apple silicon and on-device intelligence. This includes designing architectures that distill effectively, studying how teacher and student models should be trained together, and developing distillation methods that preserve reasoning, multimodal understanding, instruction following, and other important capabilities under constrained model capacity. Rather than treating deployment constraints as an afterthought, you will incorporate them into the research process—from early architecture experiments and pre-training through distillation and final model evaluation.

You may thrive in this role if you:

• Want to invent new foundation-model architectures rather than only adapt existing models.
• Enjoy combining scientific ambition with real compute, memory, latency, and energy constraints.
• Believe that small and efficient models can be a frontier research problem, not merely a compression exercise.
• Are comfortable working across model research, data, systems, and hardware boundaries.
• Care about translating research into private, useful, and deeply integrated intelligent experiences.
• Want your work to have both product impact and a presence in the broader research community.

Potential research directions include:

• Novel dense, recurrent, state-space, mixture-of-experts, and hybrid foundation-model architectures.
• Multimodal pre-training across language, images, video, audio, and sensor-derived representations.
• Compute-optimal model and data scaling, including data mixtures, curricula, tokenization, and training objectives.
• Architecture and algorithm co-design for memory-efficient and energy-efficient inference on Apple silicon.
• Offline and on-policy distillation using teacher-generated data, logits, representations, rationales, and other supervision signals.

MINIMUM QUALIFICATIONS

Hands-on experience designing, implementing, and running large-scale pre-training experiments for large language models. Experience with LLM pre-training topics such as model architecture, training objectives, data mixtures, tokenization, curricula, scaling, and optimization. Strong proficiency with modern deep learning frameworks such as PyTorch or JAX and distributed training systems. Experience evaluating pre-trained models across language understanding, reasoning, instruction following, or multimodal capabilities. Strong understanding of transformer-based architectures and current approaches to efficient or scalable foundation-model training. Master's degree, or equivalent practical experience in machine learning, computer science, or a related technical field, and a minimum of 6 years relevant experience.

PREFERRED QUALIFICATIONS

Experience contributing to major foundation-model pre-training efforts or leading architecture experiments that influence a large training run. Research contributions in model architecture, scaling laws, multimodal pre-training, optimization, efficient attention, mixture-of-experts, state-space models, or related areas. Experience with knowledge distillation, including offline or off-policy distillation, on-policy distillation, self-distillation, sequence-level distillation, logic matching, or representation transfer. Experience designing teacher-student training pipelines or transferring capabilities from large foundation models to smaller models. Experience with multimodal models spanning language, vision, video, audio, or other sensor modalities. Understanding of inference efficiency, memory hierarchy, hardware accelerators, or hardware-software co-design. Strong publication record, influential open-source contributions, or an equivalent record of applied research impact.