Perplexity has released a local version of its Computer agent for Windows PCs with NVIDIA graphics cards, letting users run multistep AI workflows directly on their own hardware without sending sensitive files to the cloud. The agent, called Portable Computer, runs a 27-billion-parameter model called Qwen 3.8 that Perplexity optimized specifically for this task. It works on NVIDIA GeForce RTX and RTX PRO workstations with 24GB or more of graphics memory, and it can analyze files, cross-reference documents, and handle recurring work entirely on the user's device.

What Makes Running AI Locally on Your Own PC Different?

For years, the promise of local AI has been straightforward: keep your data private, avoid subscription fees, and maintain control over your workflows. Perplexity's move sharpens that promise by packaging everything a user needs into a single installation. Instead of assembling a complex stack of open-source tools like Ollama, LM Studio, or llama.cpp, users get the model, the runtime, and the agent tooling all together. This matters because the local AI ecosystem has fragmented across multiple platforms and quantization formats, making setup difficult for non-technical users.

The privacy argument is the clearest selling point. A financial workflow that processes consolidated tax returns, brokerage summaries, and 1099 forms is exactly the kind of task where enterprises hesitate to send documents to a hosted chatbot. Running that work against a 27-billion-parameter model on the user's own workstation removes that objection entirely. When a task requires heavier reasoning than the local model can deliver, the agent flags it and asks the user for permission before routing anything to a cloud model.

How to Set Up and Use Perplexity's Local Agent

• Hardware Requirements: You need an NVIDIA GeForce RTX or RTX PRO workstation with at least 24GB of graphics memory; this rules out most consumer laptops and many prosumer desktops, making it a workstation-class release.
• Integrated Connectors: The agent ships with built-in connections to Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub, so it can pull data from your everyday work tools without manual setup.
• Workflow Examples: Real-world use cases include reviewing open pull requests in GitHub and flagging outdated documentation for engineering teams, parsing two years of financial records with per-page citations for tax preparation, and analyzing funnel exports to explain drops in user activation metrics.

The default model is Qwen 3.8 27B, a mid-sized open-weight model that Perplexity post-trained specifically for agent work and tuned for NVIDIA RTX GPUs. The 27-billion-parameter size represents a meaningful jump from the 7-billion-parameter models that dominated local AI a year ago. That scale allows the agent to plan across a filesystem and a set of SaaS connectors, with a clean off-ramp to a frontier cloud model for reasoning-heavy tasks that exceed the local model's ceiling.

"As local models become more capable, AI agents can handle more work directly on a PC while keeping sensitive information on the device," said Gerardo Delgado, an NVIDIA representative.

Gerardo Delgado, NVIDIA

Where Does This Fit in the Broader AI Agent Landscape?

Perplexity's move reveals a strategic split in how AI agent platforms are being built. OpenAI, Anthropic, and Google have leaned into cloud-hosted agents with browser control and hosted tool use. Perplexity is betting that a significant share of high-value agent work, particularly in legal, financial, and engineering domains where regulated data is involved, will migrate to the endpoint the moment the local model is good enough.

The timing matters. A year ago, realistic local agent work was limited to short summarization tasks with a 7-billion-parameter model. Today, a 27-billion-parameter agent can plan across a filesystem and multiple SaaS connectors, with explicit user consent required before escalating to cloud models. The gap between local and cloud reasoning is still real, but it is narrowing faster than most enterprise procurement cycles move.

NVIDIA is positioning its DGX Spark and RTX workstations as the reference hardware for this new class of local agents. In parallel, other model creators are optimizing their offerings for the same hardware. Z.ai's GLM 5.3 Flash is being optimized for DGX Station and dual DGX Spark systems, while Qwen has released Qwen 3.8-Flash-Next and an early preview of Qwen 4 that can run locally on a single DGX Spark using NVFP4 quantization. DeepSeek-v4.1 Flash is being pitched at agent workloads that need to cut key-value cache memory demands.

"Users can put the agent to work without having to research models or configure the complex software stack typically required to run local AI," noted Gerardo Delgado.

Gerardo Delgado, NVIDIA

What Are the Real Limitations?

The 24GB graphics memory floor is a significant constraint. It rules out most consumer laptops and a large share of prosumer desktops, meaning this is a workstation-class release, not something that lights up every Copilot+ PC. Support for NVIDIA DGX Station has not yet shipped. And escalation to cloud models, while gated by explicit user consent, remains the path for anything that stretches Qwen 3.8 27B's reasoning ceiling. This means the fully-local story is bounded by what a mid-sized open model can actually do.

The remaining question is whether the local model curve keeps closing the gap to frontier cloud reasoning fast enough to make on-device execution the default rather than the fallback. If local models continue improving at their current pace, the economics of keeping sensitive work on-device will become increasingly compelling for enterprises. If the gap stalls, cloud escalation will remain the norm for reasoning-heavy tasks.