PulseOnix / open source
vek
Zero-dependency, hand-tuned SIMD vector similarity kernels for dot product, L2 distance, and cosine similarity — the hot loops powering vector search, embeddings, and RAG pipelines.
By Alish Karki, part of the PulseOnix engineering collective.
Implementation
- Runtime CPU dispatch: SSE2, AVX2, AVX-512, NEON — one binary picks optimal SIMD at startup
- Multiple precisions: f32, f16, bf16, int8, uint8, and 1-bit binary vectors
- Stable C ABI — callable from C, C++, Rust, Zig, Go, Python
- Zero dependencies, stack-only hot path, cross-platform (Linux, macOS, Windows)
Usage
Integrate the C ABI into a vector-search or embedding pipeline for dot product, L2 distance, or cosine similarity. Follow the repository build instructions and select the input precision supported by your vectors; runtime dispatch selects the available SIMD path.
Limitations and evidence
vek supplies similarity kernels, not an embedding model, vector database, or approximate-nearest-neighbor index. Performance depends on CPU instruction support, vector dimensions, precision, compiler, and memory layout. Historical timings are not a portable performance guarantee.
Source and technical writing
Source code and current build instructions