2025 marked a breakout year for Artificial Intelligence development on PC, with PC-class small language models improving accuracy by nearly 2x over 2024 and the number of users downloading PC-class models growing tenfold from 2024. At CES, Nvidia is announcing Artificial Intelligence upgrades for GeForce RTX, Nvidia RTX Pro and Nvidia DGX Spark devices to support generative Artificial Intelligence on PC. The updates include up to 3x performance and 60% reduction in VRAM for video and image generative Artificial Intelligence in ComfyUI, RTX Video Super Resolution integration for faster 4K video generation, Nvidia NVFP8 optimizations for Lightricks’ LTX-2 audio-video model, a new 4K Artificial Intelligence video pipeline that integrates Blender scenes, up to 35% faster inference performance for small language models via Ollama and llama.cpp, and RTX acceleration for Nexa.ai’s Hyperlink video search.
Nvidia is introducing an RTX-powered video generation pipeline that lets artists generate videos 3x faster and upscale them to 4K while using only a fraction of the VRAM. The pipeline takes a storyboard through photorealistic keyframes to a high-quality 4K video, using a 3D object generator, a 3D-guided image generator driven by Blender, and a video generator that animates between keyframes and upscales to 4K with Nvidia RTX Video technology. It is built around Lightricks’ new LTX-2 model, which can generate up to 20 seconds of 4K video with built-in audio, multi-keyframe support and advanced conditioning features. Powered by ComfyUI, performance has been optimized by 40% on Nvidia GPUs, and with new NVFP4 and NVFP8 support, performance is 3x faster with a 60% VRAM reduction on RTX 50 Series’ NVFP4 format and 2x faster with a 40% VRAM reduction with NVFP8.
ComfyUI now offers NVFP4 and NVFP8 checkpoints for models including LTX-2, FLUX.1, FLUX.2, Qwen-Image and Z-Image, and a new RTX Video node will soon upscale clips to 4K in just seconds while sharpening edges and cleaning compression artifacts. Nvidia and ComfyUI have also improved weight streaming so system RAM can supplement VRAM for larger models on mid-range RTX GPUs, with the video generation workflow arriving next month and the LTX-2 open weights and RTX updates available now. On the productivity side, Nexa.ai’s Hyperlink uses RTX acceleration to index text and image files at 30 seconds per gigabyte and respond in three seconds on an RTX 5090 GPU, compared with an hour per gigabyte and 90 seconds on CPUs, and a new beta adds video search for objects, actions and speech. Small language model inference performance has improved by 35% for llama.cpp and 30% for Ollama over four months, with these gains coming to LM Studio and MSI’s Artificial Intelligence Robot app, while DGX Spark is receiving updates that deliver up to 2.6x faster performance since launch, along with new playbooks for speculative decoding and fine-tuning models with two Spark modules.
