This role pays less than
54%
of similar roles. Most pay
$196,750–$256,162
— the shaded band above.
At the midpoint, this role pays about
$214k
versus about
$226k
for comparable roles.
Based on 240 similar postings.
Employer
About Apple Inc
Apple Inc. is a multinational technology company known for designing and manufacturing consumer electronics, software, and online services, including the iPhone, Mac, iPad, and App Store. Industry: Consumer Electronics & Software
Apple Inc currently has
1984 open roles
on FindRole.
Listed pay typically runs
$175,000–$277,600
across 1590 roles with salary data.
Applied Researcher: On-Device Multimodal Reasoning joins the Video Computer Vision organization to develop real-time, on-device multimodal AI systems. You will design and train compact vision-language models under 10B parameters that perform multi-step reasoning over images, video, and 3D scenes while operating under strict memory and latency constraints. Your daily work involves developing inference paths including speculative decoding, KV-cache compression, and quantization, alongside distilling reasoning from frontier teachers using methods like GRPO or STaR. You will also build systems where small models invoke specialized perception experts for geometry estimation and pose recovery. The role requires expertise in Python, PyTorch, and optimization toolchains like MLX or CoreML. This position solves the technical challenge of enabling sophisticated, multi-step reasoning on mobile hardware without relying on cloud connectivity by optimizing for fixed power and performance budgets.
Design and train compact vision-language models under 10B parameters for multi-step reasoning on mobile devices.
Implement efficient reasoning techniques such as chain-of-thought distillation, adaptive compute allocation, and test-time scaling.
Optimize the decoding stack using speculative decoding, KV-cache compression, and quantization for low-latency inference.
Integrate real-time perception experts like 3D scene reconstruction and object tracking into the multimodal reasoning pipeline.
Develop structured-output fusion to represent geometry and body parameters within limited token budgets.
Apply post-training methods including supervised distillation from frontier models and reinforcement learning (GRPO, RLVR).
Drive visual token efficiency through pruning, merging, and adaptive resolution policies.
Optimize model architectures specifically for Apple silicon hardware to balance performance with power and memory constraints.
What we're looking for
Must have a Master's degree in Computer Science, Machine Learning, AI, Computer Vision, or a related field (or equivalent experience).
Must possess a strong foundation in deep learning with specific experience in LLM or VLM training and inference optimization.
Must have demonstrated experience working with multimodal models in resource-constrained environments including model compression and reasoning quality.
Must be proficient in Python and modern deep learning frameworks like PyTorch.
Must be familiar with inference and optimization toolchains such as quantization, distillation, pruning, vLLM, SGLang, llama.cpp, MLX, or CoreML.
Preferred: PhD with research in efficient multimodal reasoning, model compression, or lightweight VLM architectures.
Preferred: Experience training/post-training vision-language models end-to-end including connector design and visual instruction tuning.
Preferred: Experience deploying LLM or multimodal models on mobile or edge hardware while managing ANE/GPU kernel and memory constraints.