Donate
LAION is a non-profit. Help us build open pipelines for data, training, and evaluation, shared across mathematics and physics, biology, and multimodal world models, so anyone can study, reproduce, and improve the models that drive science.
Three pillars
Three fields are fundamental for progress on foundation models. In two of them, our open work is already published at ICLR, NeurIPS, and TMLR. Biology is next.
Models that reason through mathematics, physics, and code, built on open data recipes, benchmarks, and language-model training that anyone can reproduce.
Data recipes for reasoning models
Benchmarking agents on hard, realistic command-line tasks
Permissive-first pretraining data that front-loads math and code reasoning
Simple tasks that expose reasoning breakdowns in frontier models
Data recipes for agentic models
Open reference baselines for comparing language models and datasets
Our next pillar: foundation models for genomics, proteomics, and metabolomics that learn from DNA, proteins, and the molecules of living cells, built with the same open pipelines.
Multimodal, agentic world foundation models that see, hear, and act, as an open basis for robotics, and the methods to tell which approaches scale.
A 10-million-hour open video dataset for multimodal research
Deciding which learning procedure is more scalable
One shared engine
All three fields need the same machinery: pipelines that collect data, train models, and evaluate them. We build them openly and reuse them for every field.
We have done this before. Our OpenCLIP and CLAP pipelines took image–text and audio–text models from raw data to trained, benchmarked releases, all in the open. The next pipelines are designed so that agents can improve them too, with people reviewing every change.
Gather and curate open datasets for each field.
Train models with open, reproducible recipes.
Benchmark them openly, so anyone can check the results.
Agents propose better data, recipes, and benchmarks. Reviewed changes flow back in.
Lighthouse projects
An event where the people building open pipelines come together to push the state of the art on important models.
Every round leaves two things behind: better open pipelines, and better open datasets and models. Over time, the models that come out of Modelympics help improve the stack that produced them, so each round makes the next one stronger.
A new library for humanity. Like the library of ancient Alexandria, it should hold everything that can help train models for human progress: scientific papers with their figures and images, video, which can be scientific data too, and any other data we want to see shared at large scale.
LAION-BVD alone takes about 10 PB of storage, provided today by the Jülich Supercomputing Centre (JSC). A library for everything needs storage on a far larger scale.
Research comes first. Alexandria turns scientific texts into Knowledge Units that keep the knowledge without redistributing the original wording, and curated subsets can be made available for wider use, including in industry.
How to help
Donations pay for the researchers, engineers, and infrastructure that keep every pipeline, dataset, and model open.
Donate by bank transferGPU time powers Modelympics rounds and training across all three pillars. Tell us what you can offer and for how long.
Offer computeProject Alexandria needs durable storage far beyond the 10 PB that LAION-BVD alone takes up.
Offer storageLAION is supported by donations and public research grants. This support helps us make machine-learning datasets, models, and code available for research and public education.
Donate directly using the account details below or the QR code in your banking app. For organizational support or partnerships, see the ways to contribute.
