UForm

Why we included this project

Multimodal search often forces a tradeoff between model size and quality, but UForm tries to dodge it. Its encoder models produce Matryoshka-style embeddings from 64 to 768 dimensions for short texts and images (video is on the roadmap), which keeps vector search fast and cheap to run on modest hardware. A companion 1B chat model handles image captioning and visual question answering, so you can build lightweight in-app assistants without spinning up a giant LLaVA-style service. Because the same models export to ONNX and CoreML, you can move them from a server to a phone without re-engineering. Teams doing semantic search over visual catalogs or multilingual product listings get a practical alternative to the usual CLIP checkpoints.