Nvidia moves smart agents to local computing, speeding inference and distributing tasks across network devices

Nvidia announced at the IFA 2026 exhibition a new technology suite that aims to move the execution of AI models and autonomous agents entirely to local devices, in collaboration with Microsoft and open-source software developers, bypassing the complexities of manual installation and recurring cloud-call costs.

The move builds on the upcoming October launch of integrated RTX Spark PCs running Windows, produced by manufacturers including Lenovo with the Yoga Pro 9i and Yoga 9i models, and Acer. The devices use a chip that combines a Grace CPU with 20 cores and a Blackwell-architecture GPU delivering up to one petaflop of compute and a unified memory pool of up to 128 GB, and they integrate directly with the Windows Agent framework to run background processes under OS management.

Eliminating technical friction in running agents locally constitutes the most prominent software pillar of this announcement.Nvidia provided a one-click instant setup for the Hermes Agent application developed by Nous Research, where the system automatically detects the GPU and selects the appropriate configuration via the bundled Llama.cpp library. It also partnered with the OpenCLo community to simplify installing the task agent on RTX cards with video memory starting at 24 GB, and added support for the Perplexity Portable Computer agent that runs locally on Linux and is being prepared for Windows, enabling sensitive data to be processed on-device without consuming cloud credits, with the option to offload complex tasks to cloud models with user consent.

In terms of runtime efficiency, core improvements and speculative decoding algorithms in Llama.cpp delivered up to a 1.9× data-throughput increase on the GeForce RTX 5090, while the VLLM engine’s efficiency rose between 1.2× and 1.4× on RTX Pro 6000 Blackwell workstations and DGX Spark clusters. These enhancements support running advanced open models such as Nymtron 3.5 Lightning with 30 billion parameters, Meta’s Muse Glimmer, and the DeepSeek-V4-Flash model built with a mixture-of-experts architecture.

The release of the open-source routing tool Nvidia Peer enables the use of idle computers on the local network to distribute inference workloads.The tool automatically discovers compatible devices and distributes parallel sub-tasks for the agent, such as mailbox sorting or financial and code-report analysis, through its integration with Olama and LM Studio tools, preventing the main GPU from bottlenecking and allowing tasks to continue without affecting design or editing work on the primary machine.

This shift directly impacts work environments in the Gulf, Egypt and the broader Arab region, giving companies and development teams the ability to run agent systems and automate complex workflows locally without paying ongoing cloud subscriptions for each compute token. It allows financial, legal and startup organizations in the area to process internal documents and sensitive client data securely within their own networks, while leveraging open-source models that support local development in line with compliance and privacy requirements.

Moving agent AI to local processors and distributed routing systems marks a new phase that reduces reliance on closed cloud platforms and places advanced inference capabilities under the direct operational control of engineering and business teams.