Perplexity Portable Computer Comes to Windows: The Promise and the 24GB VRAM Catch

Perplexity is bringing a more ambitious version of local AI to Windows. Portable Computer, the company's on-device version of its Computer agent, is now available inside the Perplexity Windows app on compatible NVIDIA hardware. Instead of sending every document and reasoning step to a remote model, the agent can analyze files, use local tools and execute multi-step workflows directly on the PC.

That sounds like a major expansion of local AI, and in some ways it is. The important limitation is hardware: on-device inference requires an NVIDIA GeForce RTX or RTX PRO GPU with at least 24GB of VRAM. That requirement excludes a large share of otherwise powerful gaming and creator PCs.

The release therefore illustrates both sides of the local-agent movement. Modern consumer hardware is becoming capable enough to run genuinely useful autonomous workflows without continuously depending on a cloud service. At the same time, the memory and compute needed for those workflows still place the most capable implementations near the high end of the PC market.

Portable Computer is more than a local chatbot

The distinction between a local language model and a local agent matters.

Running a model on a PC has been possible for years. An agent adds orchestration around that model: it can break a task into steps, work across files, use tools, maintain a longer-running process and potentially schedule recurring work. Perplexity says Portable Computer brings its local agent harness, orchestrator and scheduler to Windows.

That means the useful comparison is not simply whether a local model can answer a question as well as a cloud model. The question is whether the complete system can reliably perform a workflow while preserving the advantages of local execution.

Perplexity gives examples such as analyzing private financial documents, reviewing code repositories and identifying issues in business data. Those are vendor-provided examples rather than independent performance tests, but they demonstrate the target: tasks where the source material is sensitive enough that keeping it on the machine has real value.

Privacy is the central selling point

Perplexity says work completed locally stays on the device and does not consume Perplexity Computer credits. That changes the economics and privacy model compared with a cloud-only agent.

A company could, for example, analyze an internal codebase or a folder of confidential documents without routinely transmitting the underlying files to a remote model provider. For individuals, local processing can also reduce the amount of personal information that leaves the PC.

But "local" should not be interpreted as meaning the product can never communicate with the cloud.

Portable Computer can use Perplexity Search and more capable cloud models when a task requires current information or stronger reasoning. Perplexity says the system asks for permission before sending task information off-device. That makes the approval moment important: users handling confidential material need to understand what is being transmitted, not merely whether the application itself is described as local-first.

A useful security rule is to treat every cloud escalation as a boundary crossing. Local execution reduces exposure only when sensitive context actually remains local.

The 24GB VRAM requirement is a serious filter

NVIDIA and Perplexity specify at least 24GB of video memory for on-device inference. That is not a small requirement hidden in the fine print.

Many mainstream and even high-end RTX systems have less than 24GB of VRAM. A user can own a fast gaming PC, see the RTX badge and still fail the requirement. Independent hardware coverage has highlighted how dramatically this narrows the compatible consumer hardware pool.

This is a reminder that AI-PC marketing and actual local-model requirements are different things. An NPU or modern GPU can accelerate some AI features, but a long-running agent operating a substantial model may require much more memory than lightweight image effects, transcription or summarization.

Before buying hardware specifically for Portable Computer, users should verify the exact GPU and VRAM capacity rather than assuming that a recent generation is automatically compatible.

Local agents can change the cost model

Cloud AI is usually priced around subscriptions, credits, tokens or a combination of them. Local inference shifts more of the cost upfront.

The user pays for the GPU, electricity and machine, but repeated local tasks do not incur the same per-request inference charge. Perplexity specifically says locally completed work does not consume Computer credits.

That can become attractive for repetitive workloads. A workstation that already exists for 3D rendering, video production, engineering or AI development may have unused compute capacity that can perform agent tasks without adding a new usage bill every time.

The calculation is different if someone is buying an expensive GPU solely to avoid cloud fees. Hardware depreciates, consumes power and may become insufficient as local models grow. The economically sensible choice depends on task volume, privacy requirements and whether the GPU already serves other work.

Hybrid AI may be more practical than local-only AI

Portable Computer's design points toward a hybrid future rather than a clean replacement of cloud models.

Local models have advantages in privacy, predictable availability and marginal cost. Cloud models have advantages in raw capability, continuously updated infrastructure and access to much larger compute budgets. An agent that can decide which tasks remain local and which require cloud assistance can combine both.

The difficult part is making that boundary understandable. If a user chooses local execution because a document is confidential, an automatic cloud escalation would defeat the purpose. Perplexity's permission step is therefore not a minor interface detail; it is part of the product's security model.

Organizations evaluating this kind of architecture should test exactly what information is included in a proposed cloud handoff and whether administrators can define stronger policies for sensitive data.

What Windows users should evaluate

The first question is hardware compatibility. Check VRAM, not just GPU generation or RTX branding.

The second is workflow fit. Local agents make the most sense when tasks involve private local data, repeat frequently or would otherwise generate significant cloud usage. Simple web questions do not necessarily benefit from running an agent on a large GPU.

The third is verification. An agent operating locally can still make mistakes. Keeping data on the PC reduces one class of privacy risk, but it does not guarantee that a generated analysis, code change or automated action is correct.

For consequential workflows, users should keep review and approval between the agent's recommendation and an irreversible action.

Bottom line

Perplexity Portable Computer for Windows is a meaningful step toward agents that use the power already sitting inside high-end PCs. Its local-first design can improve privacy and reduce usage-based cloud costs for the right workloads.

The 24GB VRAM floor, however, keeps this from being an ordinary Windows feature. Today it is primarily a tool for users with unusually capable NVIDIA hardware.

That limitation may become less important as local models get smaller and hardware memory grows. For now, Portable Computer is best understood as an early example of where PC-based agents are heading: more autonomous, more private and increasingly hybrid—but still demanding enough that hardware matters.