Generative AI-powered laptops and PCs are unlocking advancements in gaming, content creation, productivity and development. Today, over 600 Windows apps and games are already running AI locally on more than 100 million GeForce RTX AI PCs worldwide, delivering fast, reliable and low-latency performance.

At Microsoft Ignite, NVIDIA and Microsoft announced tools to help Windows developers quickly build and optimize AI-powered apps on RTX AI PCs, making local AI more accessible. These new tools enable application and game developers to harness powerful RTX GPUs to accelerate complex AI workflows for applications such as AI agents, app assistants and digital humans.

RTX AI PCs Power Digital Humans With Multimodal Small Language Models

NVIDIA ACE is a suite of digital human technologies that brings life to agents, assistants and avatars. To achieve a higher level of understanding so that they can respond with greater context-awareness, digital humans must be able to visually perceive the world like humans do.

Enhancing digital human interactions with greater realism demands technology that enables perception and understanding of their surroundings with greater nuance. To achieve this, NVIDIA developed multimodal small language models that can process both text and imagery, excel in role-playing and are optimized for rapid response times.

The NVIDIA Nemovision-4B-Instruct model, soon to be available, uses the latest NVIDIA VILA and NVIDIA NeMo framework for distilling, pruning and quantizing to become small enough to perform on RTX GPUs with the accuracy developers need.

The model enables digital humans to understand visual imagery in the real world and on the screen to deliver relevant responses. Multimodality serves as the foundation for agentic workflows and offers a sneak peek into a future where digital humans can reason and take action with minimal assistance from a user.

NVIDIA is also introducing  the Mistral NeMo Minitron 128k Instruct family, a suite of large-context small language models designed for optimized, efficient digital human interactions, coming soon. Available in 8B-, 4B- and 2B-parameter versions, these models offer flexible options for balancing speed, memory usage and accuracy on RTX AI PCs. They can handle large datasets in a single pass, eliminating the need for data segmentation and reassembly. Built in the GGUF format, these models enhance efficiency on low-power devices and support compatibility with multiple programming languages.

Turbocharge Gen AI With NVIDIA TensorRT Model Optimizer for Windows 

When bringing models to PC environments, developers face the challenge of limited memory and compute resources for running AI locally. And they want to make models available to as many people as possible, with minimal accuracy loss.

Today, NVIDIA announced updates to NVIDIA TensorRT Model Optimizer (ModelOpt) to offer Windows developers an improved way to optimize models for ONNX Runtime deployment.

With the latest updates, TensorRT ModelOpt enables models to be optimized into an ONNX checkpoint for deploying the model within ONNX runtime environments — using GPU execution providers such as CUDA, TensorRT and DirectML.

TensorRT-ModelOpt includes advanced quantization algorithms, such as INT4-Activation Aware Weight Quantization. Compared to other tools such as Olive, the new method reduces the memory footprint of the model and improves throughput performance on RTX GPUs.

During deployment, the models can have up to 2.6x reduced memory footprint compared to FP16 models. This results in faster throughput, with minimal accuracy degradation, allowing them to run on a wider range of PCs.

Learn more about how developers on Microsoft systems, from Windows RTX AI PCs to NVIDIA Blackwell-powered Azure servers, are transforming how users interact with AI on a daily basis.

Blog Article: Here