Recently, NVIDIA announced several major upgrades for local AI at IFA 2026. Three mainstream agent applications—Hermes Agent, OpenClaw, and Perplexity Portable Computer—will support one-click local deployment on Windows. Users no longer need to manually download models or adjust parameters; the program automatically detects the GPU and completes the configuration.
In terms of performance improvements, the latest optimizations in llama.cpp (a large language model inference framework) have achieved up to a 1.9x increase in inference throughput on the GeForce RTX 5090. Token throughput for the Qwen3.6-27B model increased by 50%, while the Qwen3.6-35B saw a 90% acceleration.
Meanwhile, the inference framework vLLM achieved a 20% speed boost on the RTX PRO 6000 Blackwell, with dual DGX Spark clusters reaching a 1.4x improvement.
Beyond making existing models run faster, NVIDIA has significantly lowered the barrier to entry for local AI deployment. Perplexity Portable Computer will launch on Windows and Linux platforms in September, offering one-click local deployment functionality, though it requires an NVIDIA GPU with 24GB or more of VRAM.
After completing local deployment, users can run workflows entirely locally without consuming credits. Only when necessary are certain tasks submitted to the cloud, and user permission is required before any upload.
Perplexity is just one of the three applications supporting one-click deployment in this update. Hermes Agent, developed by Nous Research, is a general-purpose agent. Once configured, it can call tools, maintain context across tasks, remember information across sessions, and accumulate reusable skills over time. OpenClaw claims to be the largest AI project on GitHub, having garnered over 380,000 stars.
NVIDIA has also opened beta testing for the NVIDIA PAIR tool. This tool can automatically discover all compatible PCs within the home LAN, routing independent inference requests to devices with idle computing power to enable multi-machine collaborative inference. Additionally, PAIR is compatible with Windows, macOS, and Linux, supporting RTX 20 series GPUs and newer.