ZML Releases LLMD Alpha — a Zig-Based, Chip-Agnostic Inference Server Across 5 Architectures

TechCrunch / ZML·medium signal

French startup ZML released LLMD (alpha) on 2026-07-08 (covered by TechCrunch), a self-contained inference server running Llama, Gemma, Qwen, and Mistral transparently across NVIDIA CUDA, AMD ROCm, Google TPU, Intel oneAPI, and Apple Metal. Instead of wrapping Python around CUDA kernels like vLLM/Ollama, ZML uses Zig + MLIR + OpenXLA to compile model graphs into standalone native binaries with zero Python deps. It supports continuous batching, paged attention, tensor-parallel sharding, and prefix caching, but is explicitly not production-ready yet.