On-prem agents are now a purchasing decision: Perplexity's local Computer needs $4,800 of Nvidia
Portable Computer's harness is real engineering. The bill of materials starts at a $4,800 DGX Spark or a 24GB RTX card, which turns local agents into a hardware order.
The Engineer · Build desk
What happened
- Perplexity and Nvidia have put Computer, Perplexity's agentic assistant, on the desktop as Portable Computer.
- It runs either on an Nvidia DGX Spark under DGX OS, or on an Ubuntu ARM or x64 PC with an Nvidia RTX card holding at least 24GB of VRAM.
- The Spark, described as the easier of the two paths, sells for $4,800.
- Work starts on the device, and a stalled local model can consult a cloud model only after the user sees what would leave and approves it.
Why it matters
- What it costs cost Ten local seats on Nvidia's own box is $48,000 of capital spend before anyone writes a policy for it, and it lands on a hardware budget rather than a software renewal.
- What it limits constraint With a 24GB VRAM floor and memory prices that The New Stack expects to stay high, the size of any deployment is set by how many machines can be bought, not by how many people want the agent.
- Who has to decide decision The privacy story rests on whoever is at the keyboard clicking approve or decline, which is a judgement call to be governed rather than a network boundary an auditor can inspect.
- What it sets up precedent If the harness is the differentiator, buyers should expect more comparisons of the kind only the vendor can currently reproduce, at least until the bench is actually published.
Next story · Product · 8h ago
RTX Spark laptops from six makers run AI agents on the device, starting at $2,599Confirmed2 publishers
The engineering worth reading about here is in the harness, not the model. Perplexity says the orchestrator that maintains the agent loop is deterministic code rather than a second model: the local model proposes an action, and the orchestrator assembles context, enforces policy, and runs approved tool calls inside an OS-level sandbox that restricts processes, filesystem paths and network access [1]. If that sandbox is unavailable, the harness disables itself before it makes any tool call at all [2]. Nate Kupp, Perplexity's VP of Computer Enterprise and Infrastructure, says the harness accounted for most of the engineering work [6], and that the team had to "revisit almost everything throughout the stack" to get a smaller model planning and executing multi-step tasks [4].
Some of that work compensates for a model that cannot hold as much as its spec sheet suggests. Perplexity puts Qwen3.8-27B's context window at 260,000 tokens but says it begins to struggle past 100,000 [11], which leaves roughly 38 percent of the advertised window as the working budget [18]. Hence the small core prompt with skills loaded only as needed [11], and connectors rewritten as command-line tools rather than full Model Context Protocol definitions handed to the model [12].
The benchmark numbers deserve a second look. Holding model and hardware constant, the five-point lead over Pi on the 53-task internal bench works out to fewer than three tasks [15]. The gap that carries weight is on ParseBench-100, a 100-task set covering charts, layouts, tables, text and formatting, where Computer sits about 30 tasks ahead of Hermes and roughly 51 ahead of Pi [13][17]. Perplexity says it plans to open-source the benchmark [14]. Until that happens, this is a vendor scoring its own harness on a test it wrote.
On procurement, the two supported routes are closer together than the sticker implies. A Spark comes to no more than about 3.2 times the price of the cheapest qualifying card [21], and the card route still needs a host machine around it [19]. Neither figure is heading down: The New Stack ties the floor to RAM pricing and does not expect relief soon [23].
The escalation path is where the on-prem pitch gets tested. When the local model stalls, what comes back from the cloud is text; the cloud model never touches the device's files or tools, and the local orchestrator keeps control of execution [7]. That is a defensible design, and it also places the request to disclose at precisely the moment the user most wants the task to finish [5]. The bundled connectors for Google Drive, Gmail, Slack and GitHub already send web searches and connector calls off the box [9], and the demo in which a local CSV analysis is posted to Slack is the clearest available picture of where the machine's edge actually falls [16].
What to watch
- Whether the Local Knowledge Work Bench and ParseBench-100 are released in a form that lets a third party rerun the 82.6% figure on its own hardware.
- Whether Perplexity drops the 24GB VRAM floor or adds non-RTX accelerators, which would change the entry price more than any software update.
- Movement in DGX Spark and used 24GB card pricing, since the whole deployment maths sits on top of it.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+20
- Incentives70
- Confidence60
Perspective Coverage
3 publishers- Builder
- Builder 33%
- Operator
- Operator 40%
- Investor
- Investor 27%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Perplexity says the orchestrator that maintains the agent loop is deterministic code and not another AI model: the local model proposes an action while the orchestrator assembles the context, enforces policy, and runs approved tool calls in an OS-level sandbox that restricts processes, filesystem paths and network access.
- [2]
If the sandbox isn't available, the harness disables itself before making any tool calls instead of running them outside the sandbox.
- [3]
Perplexity, in partnership with Nvidia, has taken Computer, its agentic AI assistant, to the desktop in the form of Portable Computer.
- [4]
Nate Kupp, Perplexity's vice president of Computer Enterprise and Infrastructure, told The New Stack the team had to 'revisit almost everything throughout the stack'; the company reused many of Computer's capabilities but changed the harness and model configuration for local hardware, with the model still required to plan tasks, call tools, manage files and execute multi-step tasks using fewer parameters.
- [5]
Each task starts on the local device; if the local model can't complete a step it can ask a cloud model for advice, and Perplexity says the harness selects the relevant context, flags potentially sensitive information, shows the user what would leave the machine, and asks for approval before making that call.
- [6]
Kupp says the harness accounted for most of the engineering work.
- [7]
The cloud model returns text guidance but does not get direct access to the device's files or tools, and the local orchestrator retains control of execution.
- [8]
With the same Qwen3.8-27B base model and DGX Spark hardware, Computer scored 82.6% on Perplexity's internal 53-task Local Knowledge Work Bench, compared with 77.6% for Pi and 74% for Hermes.
- [9]
Portable Computer includes connectors for Google Drive, Gmail, Slack and GitHub; when used, web searches and connector calls leave the device.
- [10]
An older RTX 3090 card with 24GB of VRAM currently costs well over $1,500.
- [11]
Perplexity says Qwen3.8-27B supports a 260,000-token context window but begins to struggle beyond 100,000 tokens, so Portable Computer keeps the core prompt and tool set small and only loads additional skills as needed.
- [12]
Perplexity turned commonly used connectors into command-line tools rather than exposing their larger Model Context Protocol definitions directly to the model.
- [13]
On ParseBench-100, a 100-task subset covering charts, layouts, tables, text and formatting, Computer scored 65.1%, Hermes 34.6% and Pi 13.9%.
- [14]
Perplexity says it plans to open-source the internal benchmark.
- [15]
Computer's 5.0-point margin over Pi on the 53-task Local Knowledge Work Bench corresponds to about 2.7 tasks.
- [16]
In one demo Portable Computer reviewed a folder of tax documents on an Nvidia DGX Spark; in a second, the local agent analyzed a local CSV file and then posted its findings to Slack.
- [17]
On the 100-task ParseBench-100 set, Computer is 30.5 tasks ahead of Hermes and 51.2 tasks ahead of Pi.
- [18]
The 100,000-token practical limit is about 38 percent of the model's advertised 260,000-token context window.
- [19]
There are two main options for running Portable Computer: a DGX Spark desktop from Nvidia running DGX OS, or a more traditional PC running Ubuntu on ARM or x64 hardware with an Nvidia RTX card with at least 24GB of GPU VRAM.
- [20]
A DGX Spark costs $4,800.
- [21]
A $4,800 DGX Spark costs no more than about 3.2 times the price of a qualifying 24GB RTX 3090, since the card is stated as well over $1,500.
- [22]
Ten seats on the DGX Spark path is $48,000 of hardware spend.
- [23]
The New Stack attributes the price level to 'RAMmageddon' and says this won't likely get cheaper anytime soon.
Sources
3 independent publishers whose own reporting we read for this story.
- dev.toPerplexity launches on-device AI for data privacy
1 article · August 27, 2026
- runtimewire.comPerplexity launches Portable Computer on DGX Spark to cut cloud use
2 articles · August 26, 2026
- thenewstack.ioPerplexity’s Computer agent can now run locally — if you can afford it
3 articles · August 26, 2026
Topics and entities
Entities
- Model Context Protocol
- Nvidia
- Perplexity Portable Computer
- Aravind Srinivas
- pi
- Perplexity
- Local Knowledge Work Bench
- Qwen3.8-27B
- NVIDIA DGX Spark