Abstract:
NVIDIA recently officially launched the Personal AI Router (PAIR) beta version during IFA 2026. This is a free open source software. It is not a network hardware router in the traditional sense, but a "virtual router" for local AI inference. It can automatically discover compatible computers in the same LAN, organize the idle computing power of multiple devices, and provide more sufficient inference resources for AI agents and local large model tasks.
Nvidia said that PAIR can dynamically adapt to the device status. When a new computer joins the network or an existing device leaves, the system can automatically adjust; the routing process will comprehensively consider factors such as node readiness, inference engine status, whether the model already exists, and GPU utilization, and then select the appropriate machine to handle the request. The entire process does not require building a server cluster in the traditional sense, nor does it require dedicated cables, racks or complex cluster configurations.
In terms of security, devices joining the PAIR cluster first need to complete pairing through a six-digit pairing code, and then an encrypted communication channel is established between the devices through two-way TLS (mTLS) to ensure that the connection between participating devices is protected. NVIDIA emphasized that one of the design goals of PAIR is to keep the local AI workflow in the user's own network, and prompt words, files, and AI agent context do not need to be sent to the cloud inference service.
Currently, the PAIR beta version supports Windows, macOS and Linux. In terms of hardware, compatibility includes NVIDIA GeForce RTX 20 series and newer models, NVIDIA RTX PRO workstation GPUs based on Turing architecture and newer architecture, NVIDIA DGX Spark, and Mac devices equipped with Apple M4 or newer chips. The system requirements given by NVIDIA also include at least 8GB of memory, and it is recommended to prepare 20GB or more disk space; the software itself does not require an Internet connection to run, but an Internet connection is still required when downloading the model.
The demonstration given by NVIDIA shows that PAIR is particularly suitable for multi-agent workflows. For example, users can ask Hermes Agent to develop a "Sunday reset" plan to sort out the messy mailbox and determine which emails need to be processed immediately, which ones can be processed later, and which ones can be skipped directly. The agent can split tasks among multiple subagents, and PAIR distributes the independent inference requests generated by these subagents to different computers in the home, rather than having all tasks compete for the same GPU. In this way, multiple tasks can be performed in parallel, and the main computer can also transfer part of the AI calculations to other devices during gaming, creation, or other work.
A five-child agent demonstration announced by NVIDIA also reflects the effect of this approach. Using Hermes Desktop and Ollama, a PAIR cluster consisting of three devices took 8 minutes and 48 seconds to complete the task, while a single RTX Spark notebook took 18 minutes. NVIDIA believes that this shows that utilizing the idle computing power of multiple devices in the home network can significantly improve the throughput of local multi-tasking AI.
An important judgment behind PAIR is that home users often already own multiple computers with certain computing capabilities, but these devices are idle most of the time. Data cited by Nvidia points out that more than half of households in the United States own two or more computers, so a large number of GPUs and other computing resources are not fully utilized for a considerable part of the day. For families with gaming computers, laptops, and other high-performance devices, PAIR attempts to transform these originally independent devices into a pool of computing resources that can jointly serve local AI.
The NVIDIA technical team even estimates that in an extreme case, there may be approximately 165 TFLOPS of idle computing power within a household. The relevant person in charge described these local computing power that has not been fully utilized as “idle free token resources at home.” However, NVIDIA expects that typical users who are truly suitable for PAIR do not need to have such an exaggerated device size. A more realistic configuration may be a MacBook or Windows notebook, plus a gaming PC.
In addition to PAIR, NVIDIA also announced this time to further simplify the deployment experience of local AI agents. Three AI agent software, including Perplexity Portable Computer, Hermes Agent and OpenClaw, will provide simpler local configuration methods for NVIDIA GPUs. In the past, users often needed to select models, find compatible inference servers, adjust quantization parameters and complete various configurations. The new solution hopes to compress this process into a few clicks.
Among them, Perplexity Portable Computer has previously been able to run on NVIDIA RTX GPUs equipped with at least 24GB of video memory and DGX Spark on Linux systems, and the Windows version is also in progress. It allows users to keep more of their workflows on-premises and choose to use cloud models when additional search or stronger inference capabilities are needed; the system will obtain user consent before handing content to the cloud.
Hermes Agent, developed by Nous Research, is a general-purpose AI agent. The local configuration of the Windows version will support one-click model deployment. The software can automatically identify the NVIDIA GPU, select the appropriate model and configuration according to the hardware, and run it through the integrated NVIDIA optimized llama.cpp. Linux support will be launched at a later date. After running, Hermes can call tools, maintain context across tasks, remember information from different sessions, and accumulate reusable skills with use.
OpenClaw also gets a simplified deployment solution for Windows. Nvidia said that OpenClaw has become an important project in the open AI agent ecosystem, and the project scale on GitHub has exceeded 380,000 stars. The new Windows application will lower the threshold for local model configuration, making it easier for users with at least 24GB of video memory NVIDIA RTX GPU to launch local AI assistants.
In order to further improve the response speed of the local AI agent, NVIDIA also simultaneously announced inference optimization for llama.cpp and vLLM. Based on new kernel optimization, more advanced speculative decoding, and faster prefill processing, the throughput of llama.cpp on GeForce RTX 5090 can be increased up to 1.9 times; vLLM performance can be improved up to 1.2 times on RTX PRO 6000 Blackwell Workstation Edition, and up to 1.4 times on a cluster composed of two DGX Sparks. These optimizations are already available through the corresponding inference backends and are also available in LM Studio and Ollama.
Nvidia also revealed this time that new Windows small AI computers equipped with the RTX Spark platform are expected to be launched in October. As these local AI-oriented hardware, inference software, and AI agent tools gradually mature, NVIDIA is clearly promoting a different way of using traditional cloud AI: users do not need to submit all tasks to remote data centers, but can directly use GPUs and other devices at home to build their own local AI computing environment.
PAIR is still in the Beta stage, and the project has been released as open source. Developers can view the source code, submit issues and participate in improvements. NVIDIA hopes that through PAIR, local AI will gradually move from "a single high-performance computer running a model" to "multiple home devices sharing the AI workload", and this may also become a new computing model after the further popularization of local AI agents.
Comments