Developer builds 15MB .NET 10 LLM engine that bypasses CUDA toolkits entirely

A developer has built Glacier.Inference, a local large language model inference engine written in pure C# using .NET 10's Native AOT compilation, producing a self-contained executable of roughly 15MB. The engine bypasses conventional CUDA runtime libraries like cudart64.dll by communicating directly with the base Windows GPU driver via low-level P/Invoke calls, replacing a toolchain that typically exceeds 4.5GB. Glacier.Inference supports NVIDIA, AMD, and Intel GPUs and reads GGUF model files using zero-copy memory mapping with cold load times under 30 milliseconds. A key optimization fuses the token selection step directly on the GPU, transferring just 4 bytes per token instead of the roughly 608KB logit tensor typically sent back to the CPU over PCIe. The project challenges the assumption that high-performance local LLM inference requires C++ runtimes or large native binary dependencies.