The Rise of the AI-Powered PC

2025 proved to be a pivotal year for AI development on personal computers. Small Language Models (SLMs) witnessed a near doubling in accuracy compared to 2024, significantly narrowing the performance gap with their larger, cloud-based counterparts. This surge in capability is fueled by maturing developer tools like Ollama, ComfyUI, llama.cpp, and Unsloth, which have seen a dramatic increase in popularity – user downloads of PC-class models soared tenfold year-over-year. This momentum is paving the way for widespread adoption of generative AI across a diverse range of applications.

NVIDIA’s AI Arsenal: Unlocking Performance

NVIDIA’s CES announcements center around unlocking the full potential of generative AI on PC hardware. Key improvements include:

• Accelerated Generative AI: Up to 3x performance gains and a 60% reduction in VRAM usage for video and image generation, achieved through PyTorch-CUDA optimizations and native NVFP4/FP8 precision support within ComfyUI.
• 4K Video Generation: Integration of RTX Video Super Resolution into ComfyUI, dramatically accelerating the creation of 4K video content.
• LTX-2 Optimization: NVIDIA NVFP8 optimizations for the open-weights release of Lightricks’ cutting-edge LTX-2 audio-video generation model, enhancing its performance and efficiency.
• AI-Powered Video Pipeline: A novel video generation pipeline leveraging a 3D scene in Blender for precise control over outputs, enabling the creation of high-quality 4K AI videos.
• SLM Speed Boost: Up to 35% faster inference performance for SLMs using Ollama and llama.cpp, streamlining AI-driven tasks.
• Intelligent Search: RTX acceleration for Nexa.ai’s Hyperlink, a revolutionary local search agent capable of understanding and responding to queries about PC files and videos.

Generating 4K Video 3x Faster: A Game Changer for Creators

Traditionally, generating high-resolution video with AI has been a significant challenge, often limited by VRAM constraints and the complexity of cloud-based solutions. NVIDIA is addressing this head-on with a new RTX-powered video generation pipeline. This pipeline empowers artists with unprecedented control over their creations, enabling them to generate videos three times faster and upscale them to stunning 4K resolution – all while minimizing VRAM usage.

The workflow is structured around three key blueprints:

• 3D Object Generator: Creates assets for immersive scenes.
• 3D-Guided Image Generator: Allows users to design scenes in Blender and generate photorealistic keyframes.
• Video Generator: Animates videos based on start and end keyframes, utilizing NVIDIA RTX Video technology for 4K upscaling.

At the heart of this innovation lies the groundbreaking LTX-2 model from Lightricks. Available for download today, LTX-2 delivers cloud-quality results locally, generating up to 20 seconds of 4K video with exceptional visual fidelity, complete with built-in audio and advanced conditioning capabilities. ComfyUI plays a crucial role, having received a 40% performance boost from NVIDIA optimizations and now supporting NVFP4 and NVFP8 data formats. This translates to 3x faster performance and 60% reduced VRAM usage with RTX 50 Series’ NVFP4 format, and 2x faster performance with a 40% VRAM reduction using NVFP8.

NVFP4 and NVFP8 checkpoints are readily available for top models within ComfyUI, including LTX-2, FLUX.1 and FLUX.2, and Qwen-Image and Z-Image. Additional model support is on the horizon.

The final touch is RTX Video, which will be integrated into ComfyUI next month, enabling real-time 4K upscaling with enhanced sharpness and reduced compression artifacts. NVIDIA has also improved ComfyUI’s weight streaming feature, allowing it to leverage system RAM when VRAM is exhausted, expanding the possibilities for larger models and complex workflows on mid-range RTX GPUs.

Hyperlink: The AI-Powered Search for Your Digital Life

For decades, PC file searching has remained frustratingly primitive, relying on filenames and incomplete metadata. Hyperlink, Nexa.ai’s local search agent, is poised to revolutionize this experience. By turning RTX PCs into searchable knowledge bases, Hyperlink allows users to ask questions in natural language and receive answers with inline citations, drawing from documents, slides, PDFs, and images. All data processing occurs locally, ensuring privacy and security. Powered by RTX acceleration, Hyperlink indexes text and image files at 30 seconds per gigabyte and responds to queries in just three seconds on an RTX 5090 GPU – a stark contrast to the hour per gigabyte indexing and 90-second response times of CPUs.

Nexa.ai is previewing a new beta version of Hyperlink at CES that extends its capabilities to video content, enabling users to search videos for specific objects, actions, and speech. This feature will be invaluable for video editors, gamers, and anyone seeking to quickly locate specific moments within their video library.

Interested users can sign up for the Hyperlink private beta on this webpage, with access rolling out this month.

Faster Small Language Models for Enhanced Productivity

NVIDIA’s collaboration with the open-source community has yielded significant performance gains for SLMs on RTX GPUs and the NVIDIA DGX Spark desktop supercomputer, utilizing Llama.cpp and Ollama. SLM inference performance has improved by 35% for llama.cpp and 30% for Ollama over the past four months. These updates are available now, and llama.cpp also benefits from faster LLM loading times. These improvements will soon be available in LM Studio and agentic apps like the MSI AI Robot app, which leverages Llama.cpp optimizations to control MSI device settings.

Broadcast 2.1: Enhanced Streaming and Conferencing

The NVIDIA Broadcast app, renowned for its AI-powered microphone and webcam enhancements, has received a significant update. Version 2.1 brings the Virtual Key Light effect to a wider range of RTX GPUs – now compatible with RTX 3060 and higher – while improving performance, handling more lighting conditions, offering broader color temperature control, and utilizing an updated HDRi base map for a professional two-key-light style. The updated NVIDIA Broadcast app is available for download today: NVIDIA Broadcast.

DGX Spark: The AI Developer’s Powerhouse

As AI models continue to evolve, the demand for powerful and flexible local AI setups is growing. DGX Spark, a compact AI supercomputer, provides developers with the resources to experiment, prototype, and run advanced AI workloads alongside their existing PCs. Ideal for LLM testing, agentic workflow prototyping, and parallel asset generation, DGX Spark is a game-changer for creative professionals.

NVIDIA is unveiling major AI performance updates to DGX Spark at CES, delivering up to 2.6x faster performance since its launch just three months ago. New DGX Spark playbooks are also available, including those for speculative decoding and fine-tuning models with two DGX Spark modules.

Frequently Asked Questions About NVIDIA’s AI Advancements

• What are Small Language Models (SLMs) and why are they important?
  SLMs are AI models that require less computational power than larger models, making them ideal for running on personal computers. Their increasing accuracy is bringing advanced AI capabilities to a wider audience.
• How does NVIDIA RTX Video Super Resolution improve video generation?
  RTX Video Super Resolution accelerates the creation of 4K video content by intelligently upscaling lower-resolution footage, resulting in sharper, more detailed visuals.
• What is the benefit of using the LTX-2 model for video generation?
  The LTX-2 model delivers cloud-quality video generation results locally, offering cinematic-level quality and control without relying on cloud dependencies.
• What is Hyperlink and how does it enhance PC search capabilities?
  Hyperlink is an AI-powered search agent that transforms your PC into a searchable knowledge base, allowing you to find files and information using natural language queries.
• What is NVIDIA DGX Spark and who is it designed for?
  DGX Spark is a compact AI supercomputer designed for developers and researchers who need a powerful local platform for experimenting with and deploying advanced AI models.
• Will NVIDIA Broadcast 2.1 work with my existing RTX graphics card?
  NVIDIA Broadcast 2.1 is compatible with RTX 3060 and higher, bringing enhanced streaming and conferencing features to a broader range of users.