Hipfire is an RDNA-native inference engine targeting the entire AMD RDNA family from RDNA1 through RDNA4 on consumer, pro, and APU hardware. It runs LLM inference without Python or the full ROCm userspace stack, instead shipping pre-compiled kernels and JIT-compiling the rest through HIP. Supports quantization, speculative decoding via DFlash, and provides an Ollama-style CLI with an OpenAI-compatible API daemon.