You'll build and train large-scale multimodal agentic models - systems that reason, plan, code, and call tools to do complex, multi-step work over pixels. This is core research shaping how users interact with what Luma's models can do.
It's a multi-stack research role across modeling, data, systems, and evaluation, on novel problems with no existing playbook, treating science and engineering as equally important. It fits someone grounded in foundation models and agentic systems who's trained models at real scale. If you want to work in only one layer of the stack, this deliberately spans several.
What You'll Own- Architect large-scale multimodal agentic models that use reasoning, planning, coding, and tool calling for complex, multi-step work.
- Design, build, and run robust data pipelines to construct, enrich, and filter massive pixel datasets, and formulate new tasks.
- Train large-scale multimodal models on massive datasets and GPU clusters.
- Define and build novel evaluation frameworks to measure multimodal agents.
First 90 DaysOne way the first 90 could unfold.
- Days 1-30 - Immerse & Diagnose: Learn the current models, agentic approaches, and where evaluation and data are weakest.
- Days 30-60 - Ship & Validate: Improve an agentic capability (reasoning, tool use, or coding) and prove it with a new eval.
- Days 60-90 - Scale & Systemize: Scale the approach across datasets and clusters and harden the evaluation framework.
What You Bring- Strong foundation in machine learning, foundation models, and agentic systems.
- Deep understanding of agentic systems and LLM/VLM reasoning, coding models, and tool calling.
- Hands-on PyTorch and large-scale training (distributed, mixed precision, large datasets).
Nice to Have- Experience with state-of-the-art foundation models in reasoning, coding, or tool calling, or state-of-the-art multimodal agents.
