Perplexity brings cloud-free agents to DGX Spark

The Portable Computer is a fully local version of the existing Perplexity Computer. According to the company, the entire runtime — the orchestrator LLM that divides and assigns work, the sub-agent LLM that actually performs the tasks, and the agent harness that connects the two to tools and files — runs on the user's own hardware. There's no cloud dependency at all, the company emphasized.

The term "harness" might sound unfamiliar. It refers to the wrapper program that lets a model do more than just spit out answers — it lets the model open files, execute commands, and read the results back. Whether an agent can actually get a job done depends as much on this wrapper as it does on the model itself.

Same model, different harness

Perplexity's own Local Knowledge Work Bench results make that point clearly. All four combinations ran on DGX Spark, and each bar carries a 95% confidence interval.

Harness + Model	Score
Computer + PPLX 27B	85.4%	85
Computer + Qwen 3.8 27B	82.6%	83
Pi + Qwen 3.8 27B	77.6%	78
Hermes + Qwen 3.8 27B	74.0%	74

The bottom three rows all use the same model: Qwen 3.8 27B. Yet swapping in Perplexity's own Computer harness moves the score from 74.0% to 82.6% — an 8.6-percentage-point jump. That's a bigger swing than what came from changing the model itself. Of course, this is Perplexity's own benchmark evaluating its own harness, so it's worth reading with that caveat in mind.

How to try it

① What you need — A DGX Spark unit and a Perplexity Pro or Max subscription. According to TestingCatalog's writeup, only these two subscription tiers can access the Portable Computer for now.

② The process — Here's what's been confirmed:

1. Install Perplexity's Portable Computer runtime on DGX Spark. The orchestrator, sub-agent, and harness all come bundled together on the device.
2. Choose which model to run. Right now the options are PPLX 27B and Qwen 3.8 27B, both fine-tuned by Perplexity, with NVIDIA's Nemotron 3.5 Lightning reportedly coming soon.
3. Assign tasks the same way you would with the regular Perplexity Computer — except now, requests and intermediate outputs never leave the device.

③ What it's suited for — Think of feeding it an entire folder of contracts to sort clause by clause, drafting a report using only internal materials, or summarizing an unreleased proposal. These are exactly the kinds of tasks that get blocked the moment data has to leave for an external server.

Riding August's DGX Spark momentum

This announcement didn't come out of nowhere. On August 11, dev-automation startup Factory unveiled a coding environment built on DGX Spark and Nemotron 3.5 Lightning that kept source code from ever leaving the device. On August 22, "Maca," a local-first agent workspace incubated by the Apache Software Foundation, hit GitHub Trending. Perplexity's move lands right in the middle of a month full of attempts to run full agents on desktop-scale machines.

The money trail overlaps too. Just a day earlier, on August 24, reports emerged that NVIDIA was in talks to invest in Perplexity at a valuation north of $30 billion. And Perplexity Computer itself has been cited as a key driver behind the company's annualized revenue jumping from $250 million to over $750 million.

Editor's take

The most valuable number in this announcement isn't 85.4% — it's the 8.6-percentage-point gap between 74.0% and 82.6%. That gap opened up with the exact same model, just by swapping the wrapper around it. This is precisely the wall that teams working with local agents have kept hitting for the past two years. Getting an open model downloaded and running on a device takes half a day. Wiring that model to open files, execute commands, and recover from failures — that's what eats weeks. What Perplexity is really selling isn't a 27B model. It's that wiring.

Anyone who's put a 27B-class model to real work tends to reach the same conclusion. For single-shot Q&A tasks, the gap with large cloud models is still quite noticeable. But for office work that repeats within a fixed folder structure and fixed format, a lot of that gap disappears. And that's exactly the kind of task the Portable Computer is targeting — "local knowledge work," true to its name. The pitch isn't that a 27B model can match GPT-level reasoning. It's that data which can never leave the building can now be processed inside it.

Korean teams might want to rethink how they budget for this. On-premise AI has traditionally started with a GPU server rack quote. But stacking an agent runtime on top of a few DGX Spark-class machines opens up a new option before you even get to that stage. That said, total cost — subscription fees plus hardware — and actual success rates on your own documents need to be measured against your own tasks, not a vendor's benchmark. An 85.4% score from 53 test tasks won't automatically show up when you run it against your company's contract folder. Building a repeatable test with 30 to 50 internal documents, using the same methodology, gives you a baseline you can trust when it's time to swap models later.

For NVIDIA, this pairing is a clear win either way. The more useful local agents become, the more it's the machines themselves — not cloud compute hours — that get sold. Once Nemotron 3.5 Lightning is added to the mix, DGX Spark shifts from being an "experimental device for developers" to something closer to "the machine a whole department runs on." Expect other agent companies to roll out their own local versions for this hardware within weeks. Factory and Perplexity both moved in August alone — the line behind them is probably already forming.