Bench Labs is an open-source collective exploring artificial intelligence together. We design and publish tiny, efficient generative models — from text-to-image diffusion to text-to-3D — alongside lightweight reasoning benchmarks, and share everything openly on Hugging Face. Everything we release is open weights under permissive licenses (MIT / Apache-2.0).
Latest release: PixelModel v6 — our 0.2B parameter multi-modal diffusion transformer for text-to-image generation, representing our most advanced architecture yet.
We actively collaborate with open-source collectives and teams that share our vision for permissively-licensed, compact, and highly-efficient artificial intelligence.
Our valued partners include AtomixLabs, FWKV, FromZero, İvmeLabs, Simplicity, and Complexity ML.