Cosmos Reason 2, Quantized for the Edge

Today we’re releasing the first quantized version of Cosmos Reason 2, which runs efficiently on the Jetson Nano Super: embedl/Cosmos-Reason2-2B-W4A16. This release brings Cosmos Reason 2’s multimodal reasoning capabilities to the full Jetson portfolio, even the most constrained devices.

Cosmos Reason 2 is NVIDIA’s multimodal reasoning model designed to take text plus images or video as input and produce high-quality, grounded text outputs, covering everything from assistant-style Q&A and planning to video analytics and perception-heavy “Physical AI” workflows. It’s powerful, but like most modern multimodal models, it’s also memory-hungry, which can make deployment on smaller edge devices challenging without optimization.

Our approach uses INT4 weight quantization (W4) while keeping activations in FP16 (A16), striking a strong balance between performance, memory footprint, and output quality. In practice, this enables Cosmos Reason 2 to run where the original model can hit memory limits, while also delivering large speedups on higher-end Jetson systems.

On Jetson Orin Nano Super, the base Cosmos Reason 2 model can run out of memory in our setup, while our W4A16 model runs comfortably and delivers strong throughput across text, image, and video inference. On Jetson AGX Orin, the quantized model roughly doubles throughput in text-only inference compared to the baseline, and meaningfully improves end-to-end latency across multimodal workloads, turning Cosmos into something you can actually deploy at the edge with a responsive user experience.

You can try the model today on Hugging Face: embedl/Cosmos-Reason2-2B-W4A16. Leading up to NVIDIA GTC, we’ll release even more efficient Cosmos variants, including additional proprietary architectural and inference enhancements aimed at pushing the limits of edge multimodal reasoning, lower latency, lower memory, and higher practical concurrency for real deployments.