PAIR
An open-source local AI traffic scheduling tool released by NVIDIA. Acting as a virtual inference traffic scheduler, it distributes AI agent inference requests across multiple local devices on a LAN for parallel execution. It is compatible with the API formats of mainstream local inference frameworks such as Ollama and LM Studio, keeping all inference requests within a user-controlled local network.
Timeline (last 90 days)
The information in this article about PAIR is based on a single demonstration source on YouTube.
The technical details PAIR has publicly disclosed are currently limited, and real-world performance metrics such as cross-device scheduling latency, hardware compatibility, and network bandwidth bottlenecks still require further empirical validation.
PAIR's privacy claims are consistent with those of local inference frameworks such as Ollama and LM Studio, with the difference being that PAIR extends the compute boundary from a single device to the home network.
PAIR supports Windows, NVIDIA's own hardware ecosystem, and Apple devices.
PAIR requires installation on all of the user's devices; once installed, it assesses each device's currently available computing power and routes tasks to the most suitable machine accordingly.
NVIDIA introduced a piece of software called PAIR, designed to aggregate the idle computing power of multiple devices on a home network into a local AI cluster.
PAIR emphasizes complete localization, with models, agents, and task execution all carried out on the user's devices, with no data uploaded to the cloud.
PAIR has dynamic load-awareness capabilities, automatically rerouting AI tasks to other idle devices when a particular device becomes occupied (such as when gaming starts).
PAIR, a partner company in Hamburg, Germany, improved its delivery speed by 10x through the use of DevTools for Agents, going from one release per week to multiple deployments per day.
PAIR built 14 highly specialized skills, including a review code skill, a Figma validate skill, and a GitHub issue test skill.
19 more timeline events
All Facts (20)
PAIR is a Personal AI Router launched by NVIDIA, used to schedule and distribute local inference requests to multiple devices on a local network for execution.
70%UnverifiedThe technical details PAIR has publicly disclosed are currently limited, and real-world performance metrics such as cross-device scheduling latency, hardware compatibility, and network bandwidth bottlenecks still require further empirical validation.
50%UnverifiedPAIR requires installation on all of the user's devices; once installed, it assesses each device's currently available computing power and routes tasks to the most suitable machine accordingly.
50%UnverifiedPAIR supports Windows, NVIDIA's own hardware ecosystem, and Apple devices.
50%UnverifiedPAIR has dynamic load-awareness capabilities, automatically rerouting AI tasks to other idle devices when a particular device becomes occupied (such as when gaming starts).
50%UnverifiedPAIR emphasizes complete localization, with models, agents, and task execution all carried out on the user's devices, with no data uploaded to the cloud.
50%UnverifiedPAIR's privacy claims are consistent with those of local inference frameworks such as Ollama and LM Studio, with the difference being that PAIR extends the compute boundary from a single device to the home network.
50%UnverifiedPAIR, a partner company in Hamburg, Germany, improved its delivery speed by 10x through the use of DevTools for Agents, going from one release per week to multiple deployments per day.
50%UnverifiedPAIR built 14 highly specialized skills, including a review code skill, a Figma validate skill, and a GitHub issue test skill.
50%UnverifiedPAIR's GitHub issue test skill is over 120 lines long, with the third phase presetting three scenarios — Chrome DevTools MCP available, unavailable, and token-saving required — as A/B/C fallback options.
50%UnverifiedKeeping inference tasks within the local network rather than relying on cloud APIs helps meet compliance requirements for data localization in industries such as finance, healthcare, and law.
50%UnverifiedInference requests within a local network do not need to travel over the public internet, resulting in lower latency and faster responses.
50%UnverifiedPAIR reduces dependence on paid cloud inference services, converts idle local compute into productivity, and can deliver cost savings over long-term operations.
50%UnverifiedPAIR plays the role of an intermediate abstraction layer, shielding upstream users from the heterogeneity and distributed complexity of the underlying hardware, while maximizing the utilization of available compute downstream.
50%UnverifiedPAIR is a virtual inference router capable of aggregating the computing resources of multiple devices within the same local area network to form a compute pool that can be uniformly scheduled by agentic systems.
50%UnverifiedPAIR's core value is to expand available compute on your local network.
50%UnverifiedAs a router, PAIR distributes requests across multiple inference endpoints, sensing in real time the compute status, load levels, and model deployment conditions of each device, and directing inference requests to the most suitable node.
50%UnverifiedPAIR is open-sourced on GitHub under the Apache 2.0 license, and anyone can fork, modify, and customize it.
50%UnverifiedPAIR handles local inference traffic scheduling for hardware entirely under the user's control, while Switchyard selects models at the API level based on quality and cost and handles decisions that may leave the local network.
50%UnverifiedPAIR is not a VRAM pooling technology; it does not merge multiple devices into a virtual machine with larger VRAM—each request still runs on a single device and cannot load oversized models.
50%