Zinc

ZINC is a local LLM inference engine for consumer GPUs. It is a program written in Zig that runs models in the GGUF format and provides a command line tool, a browser chat interface and an OpenAI compatible API, so existing clients can point at a model running on the user's own hardware. It uses a native backend for each platform. On AMD Radeon cards it supports ROCm and Vulkan, on Intel Arc it supports Vulkan, and on Apple Silicon it uses Metal with native kernels and zero copy unified memory loading. An NVIDIA RTX CUDA backend exists and is labeled experimental. The project publishes benchmarks against llama.cpp on a Radeon AI PRO R9700, and lists model rows including Qwen and Gemma families, with ZINC reporting higher decode and prefill throughput than llama.cpp in its measured runs. Each backend's results are kept separate so users can compare the build they plan to run. ZINC is aimed at people who want to run open models locally on hardware they already own, including developers who need a local endpoint for coding tools and applications. The author, Stepan Zolotukhin, is an infrastructure engineer who writes on the site about kernel tuning, quantization tradeoffs and systems design. The source is on GitHub, and there is a Discord community. The site lists no pricing or paid plans.