Docs · Blog · Discord · YouTube · X · LinkedIn
Unum builds search, storage, networking & multimodal engines from the hardware up. They keep data in RAM, VRAM or on NVMe, run on x86, Arm & RISC-V CPUs and NVIDIA, AMD & Apple GPUs, and ship stable C ABIs with bindings for over a dozen languages.
- USearch: one proximity graph for vectors, sequences, shapes & geospatial sets.
- UStore: transactional database for blobs, documents & graphs on NVMe & GPUs.
- UCall: low-latency client-server engine on io_uring, Noise encryption & SIMD.
- UForm: distributed training & inference for multimodal shared embeddings.
Underneath sit four kernel libraries: StringZilla for strings, NumKong for mixed-precision math, ForkUnion for NUMA-aware parallelism, and SmashTable for transactional containers. The same stack, adapted to scientific & scholarly computing, lives in Unum Science.
