Loading today’s stories
Product1 distinct publisher2 min readPublished
Portable Computer runs Perplexity's agent on 27B models on your own hardware. Paid tiers only, Linux first, Windows in September, and a 24GB VRAM floor most desktops miss.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
The privacy claim and the billing claim rest on the same boundary, and that boundary is drawn per task, not per install. The connectors survive: Google Drive, Gmail, Slack and GitHub can still feed the agent [3], and when a job needs the network, only that portion consumes credits and tokens [5]. So the promise is not that nothing leaves the machine. It is that you decide when something does [2]. Perplexity's own phrasing is precise about it, saying sensitive data never leaves the device without permission [4]. Permission is carrying most of the sentence.
The hardware list reads as a single-vendor spec. The Linux path wants an Nvidia DGX Spark or another Linux box with an RTX GPU, running DGX OS or Ubuntu on ARM or x64 [8]. The Windows path, when it arrives, wants an RTX card with at least 24GB of VRAM [9]. No other silicon is offered. Whatever else local-first means here, it means Nvidia-first, and the September Windows date [7] widens the operating system pool without touching the part that actually excludes people.
On performance, ZDNet repeats the familiar argument that local models answer faster than cloud ones and that sandboxed execution reduces breach exposure, without publishing a benchmark or a comparison against Perplexity's own hosted models [14]. That gap matters more than usual, because model size here is capped by the card. The choices are Qwen 3.8 27B or PPLX 27B, Perplexity's post-trained variant, with everything on the platform post-trained for accuracy and efficiency [12]. Post-training is the only lever left once the parameter count is fixed by what fits in 24GB, and Nvidia's Nemotron 3.5 Lightning, pitched at high-volume long-running work, is still listed as coming [13].
Read the eligibility rules together and the shape of the buyer becomes clear. You need a paid Perplexity tier, Pro through Enterprise Max [6], and you need a GPU that starts around $1,500 [11], which is a purchase you make before the first task rather than a bill you receive after it [16]. ZDNet's own writer, who has an RTX card with 4GB, is short by a factor of six [10][15]. Anyone whose graphics card was bought for spreadsheets and video calls is in the same position.
Which leaves one question worth measuring before anyone re-plumbs a workflow around this: what share of your agentic work can finish without a network hop [5]. That share is the only part that is genuinely free of inference fees and genuinely confined to the device [2]. The rest is the same cloud agent you were already running, with an extra approval step in front of it.
Ranked by verification strength, evidence, and original report placement.
Perplexity unveiled Portable Computer on Tuesday, a new version of its Personal Computer agent released in February, which relies on local AI models to handle agentic requests without online overhead.
With Portable Computer, the models, the user's files and the user's work stay on the machine, and the user controls if and when the AI goes online.
Users can still connect Portable Computer to external services including Google Drive, Gmail, Slack and GitHub to incorporate cloud-based files and information.
Perplexity said in its blog post: "Perplexity Portable Computer is a local-first agent. Sensitive data therefore never leaves the device without permission, and local models carry no inference fee: the system is private and cost-effective by construction."
Running the agent locally uses no billing credits; if part of a task needs to go online, only the cloud work consumes credits and tokens.
Portable Computer is available only to paid accounts: Pro, Max, Enterprise Pro and Enterprise Max subscribers, not free users.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Vendor-stated, single publisher, unbenchmarked
Product facts are specific and internally consistent (models, platform, tiers, VRAM floor, credit model), but every one of them traces to a single publisher restating Perplexity's blog post, and the cluster's two items are the same article. No hands-on test, no benchmark, no third-party verification of the privacy boundary or the speed claim.
Shipped but tightly gated; no usage data
There is a dated release with concrete availability details, which is real adoption evidence, but nothing beyond it: no user counts, deployments, enterprise references or benchmark runs. The addressable base is narrowed twice over by paid-tier-only access and an Nvidia 24GB VRAM/DGX hardware floor that the reviewing writer's own machine misses.
Overstated relative to evidence
The framing ('game changer', faster, tighter security, lower costs) runs ahead of what is shown: no benchmarks, no independent test, and a vendor quote doing the work on privacy and cost. The article does disclose the strict gating and its author's own hardware shortfall, which pulls the gap down from severe, but the headline promise still exceeds the demonstrated evidence and observed adoption.
Vendor and hardware-vendor incentives visible, undisclosed
The substantive claims originate with Perplexity's own promotional blog post, and the product's requirements route demand to specific Nvidia hardware (DGX Spark, RTX 24GB) and an Nvidia model, with Nvidia's own product characterization quoted uncritically. Coverage is a single publisher operating in a commerce-adjacent consumer-tech format, and no conflict-of-interest, affiliate or briefing disclosure appears in either item.
Low-moderate: one publisher, duplicated
The specific, checkable product details give reasonable confidence about what was announced, but confidence in the claimed benefits is low: a single publisher appearing twice, no independent testing, and no corroborating primary documentation in the supplied material.
product
Perplexity makes "ask before the cloud" a feature, and prices it at DGX Spark2 distinct publishers
product
Perplexity bills nothing for local work, which makes the escalation prompt the whole product1 distinct publisher
invest
Nvidia's Perplexity talks move its money one layer further from its own chips1 distinct publisher
build
Base Compute hands kernel tuning to agents; the carryover claim is the unmeasured part1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · August 25, 2026