NVIDIA has announced the general availability of the DRIVE AGX Thor Developer Kit, a powerful platform designed to accelerate the development of autonomous vehicles and advanced in-vehicle AI systems. Built on the NVIDIA Blackwell GPU architecture, next-generation Arm CPUs, and the new NVIDIA DriveOS 7 software stack, the kit delivers unprecedented performance and safety compliance for edge AI in automotive applications. The DRIVE AGX Thor Developer Kit is engineered for ISO 26262 ASIL-D functional safety and ISO 21434 cybersecurity standards, making it suitable for production-grade automotive development. It enables OEMs, Tier-1 suppliers, and researchers to build, test, and deploy complex AI workloads—such as vision-language-action models, large language models, and multimodal AI—directly within the vehicle. The platform features a 64 GB LPDDR5X memory system, 256 GB UFS storage, and 273 GB/s memory bandwidth, powered by the Thor SoC. It includes 14x Arm Neoverse V3AE cores, offering up to 2.3x higher SPECrate 2017_int_base performance compared to the Orin SoC. The Blackwell GPU supports FP32, FP16, INT8, and FP4 precision, delivering up to 1,000 INT8 TFLOPS and 2,000 FP4 TFLOPS, enabling efficient execution of generative AI and large-scale deep learning models. The kit supports a wide array of automotive interfaces, including 16x GMSL2, 2x GMSL3 camera inputs, 10GbE Ethernet, and DisplayPort for high-resolution displays. It also includes PCIe Gen5 connectivity and a 350W TDP, making it ideal for both bench-level prototyping and vehicle integration. Key software innovations include NVIDIA DriveOS 7, which provides a unified, safety-certified foundation for AI-driven automotive systems. It integrates with TensorRT 10, offering advanced optimizations like INT4 weight-only quantization, block-wise scaling, NVFP4 precision for Blackwell GPUs, and tiling for better memory efficiency. The new version also introduces faster build times, improved debugging tools, and support for dynamic batching and speculative decoding. A major highlight is the DriveOS LLM SDK Runtime—a pure C++ inference engine optimized for low-latency deployment of LLMs and vision-language models directly in vehicles. It supports popular models like Llama 3 8B, Llama 3.1 8B, Qwen2.5-7B, and Qwen2-VL series, with FP16, INT4, FP8, and NVFP4 precision options. Developers can use it to enable personalized AI assistants, driver monitoring, and multimodal interactions. NVIDIA DriveOS Linux profiles offer three deployment-ready configurations: development, production, and test—each tailored for different stages of the development lifecycle, with varying levels of security, safety, and debug capabilities. The DRIVE AGX Thor Developer Kit is now available for pre-order, with delivery scheduled for September 2025. It comes in multiple SKUs to suit different development needs. Developers can get started by registering for an NVIDIA Developer account, joining the DRIVE AGX SDK Developer Program, and downloading DriveOS 7.0.3. NGC DriveOS Docker containers are available to streamline setup, and community support is accessible via the DRIVE AGX Thor Developer Forum.