UForm is an open-source family of pocket-sized multimodal AI models from Unum Cloud, built for content understanding and generation across text and images in many languages. The library has about 1,300 stars on GitHub, is licensed under Apache-2.0, and was updated in 2025.
Most multimodal models are too large for edge devices and cost-sensitive products. UForm goes the other way: tiny embedding models with dimensions from 64 to 768 and a roughly one-billion-parameter chat model, engineered to run fast on modest hardware. The project claims up to five times faster performance than OpenAI CLIP and LLaVA on comparable tasks. Models ship in ONNX, CoreML and PyTorch formats, with SDKs for Python, JavaScript and Swift, so the same small models can power semantic image search, cross-modal retrieval, clustering and lightweight visual question answering on servers, in browsers or natively on Apple devices.
The models are distributed through Hugging Face for easy loading. The generative models can caption images and answer questions about them, while the encoder models produce joint embeddings that place texts and images in the same vector space. Because the models are small, they are a popular choice for on-device retrieval, and for organizing or deduplicating large media libraries without sending data to a cloud API.
UForm is free and open source under the Apache-2.0 license. The models and SDKs are available on GitHub and model hubs at no cost; you run them on your own hardware.
UForm is for developers building semantic search, image retrieval or multimodal features who need models small and fast enough for edge devices, mobile apps or high-volume services.