NVIDIA has unveiled a series of groundbreaking advancements in generative AI technology at CES 2025, marking a pivotal moment for AI development on personal computers. The new features, which promise to enhance the performance and usability of generative AI on PC-class devices, come in response to a significant uptick in interest among creators, gamers, and productivity users. Small language models (SLMs) have seen a nearly 2x improvement in accuracy compared to 2024, helping to bridge the gap with larger, cloud-based language models.
Among the major announcements, NVIDIA highlighted enhancements to its GeForce RTX, RTX PRO, and DGX Spark devices, aiming to unlock the necessary performance and memory for developers to effectively deploy generative AI on personal computers. These upgrades include up to 3x performance improvements and a 60% reduction in VRAM usage for video and image generative AI through optimizations in PyTorch-CUDA. Additionally, the RTX Video Super Resolution feature has been integrated into ComfyUI, enabling accelerated 4K video generation.
The new video generation pipeline, powered by NVIDIA’s RTX technology, allows creators to produce videos three times faster while maintaining a high quality, upscaled to 4K. This innovation is particularly significant given the challenges associated with generating high-resolution videos directly on PCs, where model sizes often exceed available VRAM. The pipeline includes features such as a 3D object generator, a 3D-guided image generator, and a video generator that animates videos based on user-defined keyframes, allowing for greater artistic control.
The introduction of Lightricks’ LTX-2 model, now available for download, represents a major milestone for local AI video creation. This model can generate up to 20 seconds of 4K video with impressive visual fidelity, complete with built-in audio and advanced conditioning capabilities. The collaboration between NVIDIA and ComfyUI has yielded a 40% performance optimization on NVIDIA GPUs, with new support for NVFP4 and NVFP8 data formats, resulting in a substantial increase in efficiency.
A key feature of the latest software updates includes improvements to weight streaming, allowing ComfyUI to utilize system RAM when VRAM capacity is exceeded. This facilitates the use of larger models and more complex workflows on mid-range RTX GPUs, broadening access for creators working in AI.
Alongside these advancements, Nexa.ai introduced a beta version of Hyperlink, a local search agent that transforms traditional PC file searching into a more intuitive experience. By enabling natural language queries, Hyperlink allows users to search their documents, slides, and images without relying solely on file names. The new version extends this capability to video content, enabling users to search through videos for specific objects and actions, catering to a diverse user base from content creators to gamers.
In addition, NVIDIA has reported a 35% increase in inference performance for SLMs using models like Llama.cpp and Ollama, which have gained traction in the open-source community. These improvements are expected to enhance application performance in various AI-driven tasks, exemplified by the upcoming MSI AI Robot app that will leverage these optimizations to allow users to manage their device settings seamlessly.
The latest updates to the NVIDIA Broadcast app, including Version 2.1, improve the quality of microphone and webcam input for livestreaming and video conferencing, making advanced AI features like Virtual Key Light accessible to a broader range of users. This update enhances performance and expands compatibility, catering to content creators and professionals seeking to improve their streaming quality.
Finally, NVIDIA’s DGX Spark, a compact supercomputer designed for AI development, has experienced substantial performance improvements since its recent launch, now boasting up to 2.6x faster processing capabilities. This technology is tailored for those looking to experiment with large language models or develop AI workflows without compromising their primary computing resources.
As NVIDIA continues to push the envelope on local AI capabilities, these advancements not only signify a shift in how generative AI can be utilized on personal computers but also highlight the growing emphasis on privacy and performance in AI development. The trajectory for 2025 suggests a significant shift towards enabling a broader range of users to engage with and benefit from cutting-edge AI technologies.
See also
Generative AI Market Set to Reach $415.45 Billion by 2031, Driven by Text-to-Image Tech
Generative AI Expands Creative Horizons, Unlocking the Era of the “Super Subconscious”
Agents Surge in 2025: OpenAI and Anthropic Lead AI’s Evolution with New Protocols
Research Evaluates AI Definition Accuracy Using Cosine Similarity Metrics Across GPT Models
Yann LeCun Slams LLMs as ‘Dead End’ for Superintelligence, Advocates for World Models
