NVIDIA Unveils Local AI Acceleration and RTX Spark PCs at IFA 2026

At IFA 2026, NVIDIA announced new tools and hardware in collaboration with Microsoft and partners to accelerate local AI inference speeds and simplify agent setup. The core of this announcement is the improvement of local inference speeds by up to 1.9x through optimizations for llama.cpp and vLLM. These optimizations are immediately available via LM Studio and Ollama, and NVIDIA unveiled NVIDIA PAIR, a personal AI router tool that intelligently distributes inference tasks among PCs within a local network.

Simplifying Local Agent Execution Environments

To lower the barrier to entry for running agents based on local models, major agent apps such as Hermes Agent, OpenClaw, and Perplexity Portable Computer are introducing support and simplified setup for NVIDIA GPUs. In particular, Perplexity Portable Computer packages models and tools into a single app for Linux systems, with Windows support to be added soon. This tool adopts a hybrid approach that keeps sensitive information local while escalating tasks to frontier models in the cloud only when necessary.

RTX Spark and Expansion of the New Model Ecosystem

In October, NVIDIA RTX Spark Windows PCs from Lenovo and Acer will be released. Major game companies such as EA and Ubisoft are preparing titles for RTX Spark, providing an environment where developers and creators can run powerful AI agents locally. Additionally, various local AI models were released in August, further enriching the ecosystem.

• Nemotron 3.5 Lightning: A 30 billion parameter model runnable on RTX PCs and DGX Spark, among others
• GLM-5.3-Flash: Z.ai's multimodal MoE model supporting agentic AI on DGX Station
• Qwen3.8 Series: Open-weight multimodal MoE models and a 27 billion parameter model optimized for local coding and agent workloads
• LTX 2.5: An open-world video generation model optimized for NVIDIA GPUs and DGX systems
• MiniMax-H3: An audio-synchronized video generation model runnable locally via ComfyUI, with performance improved 7x through FastH3
• Meta Muse Glimmer: A 30 billion parameter open-weight model supporting coding and agent workloads
• DeepSeek v4 Flash: A 284 billion parameter MoE model runnable on a 2x DGX Spark cluster