Project Zero

Project Zero is an inference engine written in C99 that runs BitNet b1.58 ternary models and dense GGUF models on CPU, with no GPU, no Python and no ML framework. The build produces a single binary that needs only a C compiler, pthreads and libm.

It works as a CLI, an interactive REPL, an embeddable C library, or an OpenAI-compatible HTTP server with a bundled web chat UI. The server exposes /v1/chat/completions plus Prometheus metrics, OpenAPI docs and optional API-key auth.

SIMD kernels are selected at runtime across AVX-512 VNNI, AVX2 and NEON, and quantization covers BF16 through INT4, Q4_K, Q6_K and ternary. MIT licensed. Early-stage: one tagged release and a single primary maintainer, so treat it as something to experiment with rather than production infrastructure.