About Us

Huawei is a global leader in information and communications technology (ICT), driving innovation and shaping the future of an intelligent world. With a presence in over 170 countries and a strong commitment to research and development, Huawei continues to push the boundaries of technology.

Huawei's European Research Institute (ERI) brings together more than 1,500 researchers across over 20 sites in Europe, including our Research Center in Switzerland.

Our Swiss research laboratory focuses on pioneering next-generation computing architectures and systems for the AI era, spanning innovative hardware, software, and algorithms. Working closely with leading universities and academic partners worldwide, we foster an open, collaborative research environment dedicated to fundamental scientific breakthroughs and long-term innovation.

Our vision is to become a world-leading research center for computing architectures and systems in the AI era, shaping the future of intelligent computing.

About the Team

We are a research team based in Europe, working at the intersection of Generative AI, Video and Efficient AI. We collaborate closely with leading universities and research institutions, and with engineering teams, to explore and build the next generation of efficient multimodal AI systems.

Our work centres on efficient real-time video generation. We are moving toward mixture-of-experts (MoE) video models, streaming world models, and models that keep learning while they generate.

The team has two streams: Architecture and Systems, and Multimodal Efficiency. This position is in the Multimodal Efficiency stream, which makes video and multimodal models faster and cheaper at the model and algorithm level, and works closely with the systems stream to run them in real time on NPU/GPU clusters.

The Role

We are looking for a Research Scientist who has recently completed a PhD to make large multimodal and video generation models faster and more efficient. You will work hands-on on research projects within the Multimodal Efficiency stream: implementing ideas, running experiments at scale, building working systems and publishing, with mentoring from senior researchers.

What You Will Work On

Depending on your background, you will work hands-on on one or two of these topics and contribute across the others.

· Token efficiency for video: adaptive, content-aware token compression and merging, and positional encodings (e.g. RoPE) designed to work with compressed or multi-resolution tokens

· Long-horizon and streaming generation: memory- and KV-cache-efficient autoregressive video, history compression, and few-step and streaming distillation

· Sparse and structured attention for long video sequences

· Step distillation, caching across denoising steps, and pruning

· Quantization and low-precision inference (FP8/FP4, MX formats) for diffusion and flow-matching models

· MoE video models: expert routing and MoE-specific acceleration

· Efficient end-to-end execution of large multimodal models, including VAE encoding and decoding, and joint audio-video generation

· Efficient agentic multimodal pipelines: agents that call LLMs, VLMs and video/speech generators, with KV cache reuse across steps and latency-aware planning

· Workload analysis and hardware-aware model design: compute, memory and latency models that guide algorithm choices on NPUs and GPUs, together with the systems stream

Your Responsibilities

· Drive research projects within the team roadmap: from idea to prototype to results at scale

· Implement, train and optimize models yourself, and turn results into working systems and demos

· Run careful experiments: strong baselines, ablations, and measurements of quality, latency and memory

· Work with the systems stream and engineering teams to profile and accelerate models on NPU/GPU clusters

· Publish at top venues (CVPR, ICCV/ECCV, NeurIPS, ICLR, ICML, SIGGRAPH, MLSys) and contribute to open-source projects

· Propose new research ideas

· Help supervise interns as you grow in the role

What We Are Looking For

· PhD in Computer Science, Electrical Engineering, Artificial Intelligence, Machine Learning or a related field, completed recently or to be completed within 6 months

· At least one first-author publication at a top venue in generative AI, computer vision, multimodal AI or efficient ML

· Hands-on experience training or fine-tuning diffusion, flow-matching or other large generative models

· Strong programming skills in Python and PyTorch; C++ is an advantage

· Proven ability to carry a research project from idea to implementation and result

· Clear communication, in writing and in talks

Experience in one or more of the following areas is an advantage:

· Video generation, streaming or autoregressive video, or world models

· Token merging or pruning, adaptive tokenization, or positional encodings for multi-resolution inputs

· Step distillation, sparse or efficient attention, quantization

· MoE models

· KV cache optimization and long-context modeling

· Profiling and optimizing models on NPUs or GPUs

· RL or preference optimization for generative models

· Open-source AI projects

· Distributed training or inference

What We Offer

· Hands-on research on Multimodal AI, Video Generation, World Models and Efficient AI

· Mentoring by senior researchers, and a path to leading your own research topics

· Close collaboration with researchers and engineers from leading universities, research institutions and industry teams

· Access to large-scale AI models and advanced AI computing platforms

· Support for publishing, open-source contributions and conference attendance

· The opportunity to co-supervise interns

· A highly international research environment in Europe

Who We Are Looking For

We are particularly interested in researchers who are not only excited about making AI models smarter, but who want to answer:

· How can we make large multimodal models faster, more scalable and more efficient, and run them in real time on real hardware?

If you want to do hands-on research on efficient Multimodal AI and Video Generation, we would love to hear from you.