Open-weight model compilation, verified local inference, quantization, speculative decoding, model architecture and trustworthy AI.