Fucina
A CPU-first, eager-execution tensor and autograd library written in pure Zig, with compile-time-checked named tensor axes and no C/C++ or Python dependency.
Fucina is a tensor and autograd library written in pure Zig, where every operation executes immediately on real buffers rather than building a graph to compile first. Tensor axes carry names checked at compile time, so a contraction between mismatched axes is a compile error rather than a runtime shape crash, and the same code path runs for both inference and training.
It has no C or C++ build system and no Python runtime dependency, using Zig vector kernels with optional CBLAS providers and Metal or CUDA GPU offload for matrix multiplication. To prove the library on real workloads, the repository ships complete applications built on it, including LLM chat and serving for models such as Qwen3, DeepSeek, GLM, and Gemma, plus speech, vision, and audio models, each benchmarked against llama.cpp.
It requires the pinned Zig 0.16.0 toolchain and is fetched as an ordinary Zig package dependency.