Reflex

Reflex is a Rust and CUDA-based inference engine designed for fast cold-start latency on NVIDIA GPUs. It compiles CUDA kernels ahead-of-time for rapid process launch to first token generation, focusing on single-shot or edge deployment scenarios rather than sustained throughput. It supports GGUF model format for Qwen and DeepSeek model families and includes CLI subcommands for generation, candidate scoring, benchmarking, and IPC interfaces.