Perplexity just flipped a switch on something that’s been quietly building in the background for months: a version of its AI agent that runs entirely on your own hardware, no cloud required. The company calls it Portable Computer, and it’s designed to give people and teams the same multi-model, multi-tool experience as Perplexity’s cloud-based Computer—but with the data staying put on a local machine.

The timing isn’t accidental. It lands alongside NVIDIA‘s DGX Spark, a compact “personal AI supercomputer” built around the Grace Blackwell GB10 chip, with 128GB of unified memory and up to a petaFLOP of AI throughput in a desktop-friendly chassis. Perplexity’s Portable Computer is built specifically for that hardware at launch, turning what could have been a very powerful but generic AI workstation into a turnkey setup for running serious on-premises agents.

Why “local-first” matters now

For a lot of organizations, the promise of AI agents has always come with a catch: to get the best results, you send your documents, code, and internal data out to someone else’s servers. That works fine for public research or casual tasks. It gets complicated when you’re dealing with confidential contracts, unreleased product specs, or customer data that’s covered by strict compliance rules.

Portable Computer is Perplexity’s answer to that tension. The entire runtime—the orchestrator model that plans tasks, the sub-agent models that execute them, and the “harness” that ties tools and apps together—runs on the local device by default. Sensitive files never leave the machine unless the user explicitly authorizes an escalation to the cloud for heavier reasoning or broader web research, and even then, the system is designed to keep sensitive tokens on-device and only send what’s needed.

From a cost perspective, this also changes the math. Running inference locally means those tokens don’t hit per-call API bills. For teams that already own or are considering DGX Spark–class hardware, a big chunk of day-to-day agent work can happen without racking up usage charges, with the cloud reserved for the moments when it actually adds value.

What Portable Computer actually does

If you’ve used Perplexity Computer since its February launch, Portable Computer will feel familiar in terms of capabilities, just shifted onto your desk. It can still:

• Read and synthesize files—PDFs, docs, codebases, logs—and connect them to web search or internal connectors when allowed.
• Run multi-step workflows: research a topic, draft a report, generate code, test it in a sandbox, then iterate based on errors.
• Coordinate multiple models and tools under one interface, rather than making you juggle separate chat windows and APIs.

The difference is where that work happens. In the portable setup, the orchestrator LLM, sub-agent LLMs, and agent harness all live on your machine. When the local system decides it needs something beyond its on-device capacity—say, deeper reasoning or a wider web sweep—it asks the user for permission before escalating to Perplexity’s cloud models. That gate keeps the default posture private and local, while still allowing the agent to “phone home” when it’s truly useful.

Security-wise, Perplexity says Portable Computer matches the cloud version’s safeguards: code and tool execution run in isolated sandboxes with tightly controlled access to files and connected apps. Local-first deployments also tend to keep network exposure minimal by default, using loopback interfaces and authenticated tunnels for remote access instead of opening services directly to the public internet.

The hardware piece: DGX Spark as the launch platform

Portable Computer isn’t tied conceptually to one box, but its first official home is NVIDIA’s DGX Spark. That machine is small enough to sit on a desk—about 150mm square and 50mm tall—but packs a 20-core Arm CPU, a Blackwell-based GPU, 128GB of coherent unified memory, and up to 4TB of fast NVMe storage. NVIDIA positions it as a personal AI supercomputer capable of running models up to around 200 billion parameters locally, and even larger setups in dual-system configurations.

For Perplexity, that’s the sweet spot: enough memory and compute to run serious agent workloads without relying on a data center, but in a form factor that a small team or power user can actually own and operate. As of late August, Portable Computer is available to Perplexity Pro and Max subscribers (including Enterprise tiers) on DGX Spark systems running Linux, with Windows support expected to follow. The company has also signaled that support for more conventional RTX GPU PCs is coming later, which would broaden the hardware base beyond DGX.

On the model side, the initial release ships with Perplexity’s own PPLX 27B model and offers Qwen’s 27B variant as an option, with NVIDIA’s Nemotron 3.5 Lightning slated to arrive soon. That mix gives users some flexibility in how they balance speed, quality, and resource usage on their local rig.

Who this is really for

You can think of Portable Computer as aimed at three overlapping groups:

• Teams in regulated or sensitive domains—legal, finance, healthcare, enterprise R&D—who want agent-level automation but can’t freely send internal data to the cloud.
• Companies already investing in on-prem AI infrastructure, especially those eyeing or already using DGX-class hardware, who want a polished agent layer on top of their existing stack.
• Power users and small studios that generate a lot of proprietary content—code, design assets, research notes—and would rather keep their working data under their own control.

For these users, the value proposition isn’t just “AI on your machine.” It’s “AI that can act like a digital worker, but with your data boundaries enforced by default.”

Where this fits in the bigger AI story

Perplexity’s move mirrors a broader shift in the industry. Over the past year, we’ve seen a steady push toward local and edge AI: laptops and desktops with dedicated AI accelerators, frameworks optimized for on-device inference, and a growing menu of models designed to run without a data center. At the same time, the agent conversation has moved from “chat that can call tools” to “systems that can plan, execute, and iterate over complex tasks.”

Portable Computer sits right at that intersection. It takes the agent paradigm—multi-step, multi-tool, multi-model workflows—and marries it to a local-first architecture that addresses two of the biggest enterprise objections to AI: privacy and cost. By making the local path the default and the cloud path optional and user-gated, Perplexity is effectively saying: you don’t have to choose between powerful agents and control over your data.

The bet now is whether enough organizations and individuals will invest in hardware like DGX Spark—or future RTX-based setups—to make local-first agents a mainstream pattern rather than a niche. If the economics hold up and the tooling keeps improving, “your AI, on your machine” could become as normal as running a local database or version control server.