Gavin Li is the founder and CEO of Anima AI and the author of AirLLM, an open-source library for running large language models with limited GPU memory. His public professional biography also identifies previous AI leadership work at Airbnb and Alibaba.[1][2]
Start with AirLLM →
X Career
Li’s public work combines model infrastructure with creative AI products. His AirLLM repository links his technical writing and projects for game animation and image editing. His current founder role is self-reported; his professional biography describes earlier work at Airbnb and Alibaba without giving exact titles or employment dates.[1][2][3]
Contributions
AirLLM splits a model into layer files and brings one layer at a time onto the GPU. That reduces the amount of model weight data that must remain in video memory, while requiring substantial storage and repeated data transfer. Prefetching overlaps some loading with computation. The approach makes experimentation possible on smaller GPUs, but low memory use does not imply the throughput of a machine that holds the whole model at once.[3]
Ideas
The repository documents model-specific examples, compression options, and adapter training that streams frozen weights while keeping adapters resident. Checkpoint storage, loading time, and model support matter alongside GPU-memory requirements.[3]
Tools, writing & talks
Tools
Start here
AirLLM
github.comLayer-by-layer model loading that reduces GPU memory needs for inference and selected adapter-training workflows.
Begin with Quickstart and the model-specific storage requirements. Python, model-download access, and basic GPU-memory knowledge are useful prerequisites.
Creator[2][3]Original source