Native inference engine optimized for DeepSeek V4 Flash, GLM 5.2, and PRO models with Metal, CUDA, and ROCm support
