Show HN: I cross-compiled llama.cpp to run on Windows XP (okt.ai)

A developer cross‑compiled llama.cpp on macOS to run on Windows XP x64 and released a reproducible deployment package that includes 70+ executables (llama-cli, llama-bench, llama-quantize, GGUF model support) and an SSE4.2-optimized build for era‑appropriate CPUs. After a series of build-and-debug stages the author solved compatibility issues by targeting _WIN32_WINNT=0x0502, installing the VC++ 2019 Redistributable v16.7 (the last XP‑compatible UCRT), replacing Vista+ SRWLOCK threading primitives with XP‑compatible CRITICAL_SECTION/Event code, and downgrading cpp-httplib to v0.15.3 to avoid a deliberate modern Windows version check. The resulting XP VM runs local inference at ~2–8 tokens/sec on small models (≈0.5B), with a ~120 MB deployment package; server mode, GPU acceleration, and models >~3B remain unsupported. This is significant because it demonstrates llama.cpp’s portability to extremely old OSes and highlights practical engineering tradeoffs when bringing modern ML tooling to legacy environments—toolchain flags (disable AVX/AVX2/FMA, enable SSE4.2), explicit Windows target macros, careful runtime and library versioning, and manual threading fallbacks matter as much as model code. The repo includes a toolchain file, build flags, and a how‑to for others; beyond the novelty, the work is useful for embedded/legacy deployments and as a reproducible case study of human+AI assisted debugging (the author credits Claude for guidance).