Perplexity Portable Computer Brings Local AI Agents to Nvidia Hardware

Perplexity has launched Perplexity Portable Computer, a local-first version of its agentic AI platform designed to run primarily on users’ own hardware. Developed in partnership with Nvidia, the product aims to give users greater control over how their data is processed while reducing reliance on cloud infrastructure.

The launch represents an expansion of Perplexity’s Computer platform, which previously relied more heavily on cloud-based AI models for multi-step tasks.

How Perplexity Portable Computer Works

Perplexity Portable Computer runs a post-trained 27-billion-parameter AI model directly on compatible hardware. Users can also choose Qwen 3.8 27B, while support for Nvidia Nemotron 3.5 Lightning is planned.

The local system includes the AI orchestrator, subagent models and agent harness, allowing many tasks to be processed on the user’s computer rather than automatically sending them to a remote server.

However, the platform is not completely disconnected from the cloud. When a task requires more advanced reasoning than the local models can provide, the orchestrator can ask the user for explicit permission to use a cloud model.

Perplexity describes this process as “user-gated, PII-flagged, and text guidance only,” giving users a choice before information is sent to a cloud service.

Reported Local AI Performance

Perplexity says its local 27B model, when used with the company’s agent harness, achieved an 82.6% score on its real-world benchmarks.

The company also says locally executed workflows do not count toward users’ token limits.

These performance results are based on Perplexity’s own testing and should therefore be regarded as company-reported figures rather than independently verified benchmark results.

Nvidia Hardware Requirements

The first release of Perplexity Portable Computer is available on Nvidia’s DGX Spark, a desktop AI system featuring 128GB of unified memory and up to 1 petaflop of AI performance.

The platform also supports Linux computers equipped with Nvidia RTX GPUs containing at least 24GB of VRAM.

This hardware requirement means the initial version is primarily aimed at users with high-performance AI-capable computers.

Windows support is expected in September, giving the service a broader potential user base in the future.

Availability for Perplexity Subscribers

Perplexity Portable Computer is available to Perplexity Pro and Max subscribers without an additional charge beyond their existing subscription.

Enterprise Pro and Enterprise Max users can also access the product.

The requirement for compatible hardware, however, means access to the software does not necessarily mean every subscriber can immediately run it on their existing computer.

A Stronger Focus on Local AI

Privacy is one of the main reasons behind the local-first approach.

By processing many AI tasks directly on the user’s machine, Perplexity Portable Computer can reduce the amount of information that needs to be transferred to cloud infrastructure. Users retain control over whether more demanding tasks should be sent to cloud models.

The approach could be particularly useful for workflows involving sensitive documents or information where users want greater control over data processing.

At the same time, the system’s cloud escalation option means it should not be described as a completely offline AI platform.

From Computer to Portable Computer

Perplexity originally introduced its Computer platform in February 2026 as a cloud-based multi-model agent designed to handle knowledge-work tasks.

The company subsequently expanded Computer to Windows desktops through its Personal Computer software.

Perplexity Portable Computer takes a different approach by moving more of the AI workload onto the user’s own hardware.

The shift reflects a broader industry trend toward local-first AI, where increasingly capable models can run directly on personal computers and workstations rather than requiring every request to be processed in the cloud.

Local AI and Token Usage

Another notable aspect of the product is its approach to token consumption.

Perplexity says tasks completed locally do not count toward token limits associated with cloud AI usage. If users approve a cloud escalation, however, the request can involve remote AI infrastructure.

This model allows users to perform supported workloads locally while retaining access to more powerful cloud models when necessary.

Growing Competition in Local AI

The launch places Perplexity among a growing number of companies exploring local and hybrid AI systems.

Improvements in GPUs, memory capacity and smaller language models are making it increasingly practical to run sophisticated AI workloads outside traditional data centres.

Perplexity’s collaboration with Nvidia highlights the importance of specialised hardware in this transition. Systems such as DGX Spark provide the memory and computing resources required to run larger AI models locally.

Conclusion

Perplexity Portable Computer expands Perplexity’s agentic AI strategy by bringing more AI processing directly onto compatible user hardware.

The platform supports a local 27B model, Qwen 3.8 27B and planned support for Nvidia Nemotron 3.5 Lightning. It initially runs on Nvidia DGX Spark and qualifying Linux systems with RTX GPUs, while Windows support is expected in September.

The product is not entirely offline because users can approve cloud processing for tasks requiring stronger reasoning. Instead, its key distinction is a local-first approach with user-controlled cloud access.

As local AI hardware and models continue to improve, Perplexity Portable Computer demonstrates how AI agents could increasingly combine on-device processing with optional cloud capabilities.