On-prem agents are now a purchasing decision: Perplexity's local Computer needs $4,800 of Nvidia

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 ^[8]. If that sandbox is unavailable, the harness disables itself before it makes any tool call at all ^[9]. Nate Kupp, Perplexity's VP of Computer Enterprise and Infrastructure, says the harness accounted for most of the engineering work ^[7], and that the team had to "revisit almost everything throughout the stack" to get a smaller model planning and executing multi-step tasks ^[6].

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 ^[10], which leaves roughly 38 percent of the advertised window as the working budget ^[3]. Hence the small core prompt with skills loaded only as needed ^[10], and connectors rewritten as command-line tools rather than full Model Context Protocol definitions handed to the model ^[11].

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 ^[1]. 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][2]. 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 ^[4], and the card route still needs a host machine around it ^[2]. Neither figure is heading down: The New Stack ties the floor to RAM pricing and does not expect relief soon ^[5].

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 ^[16]. 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 ^[15]. The bundled connectors for Google Drive, Gmail, Slack and GitHub already send web searches and connector calls off the box ^[17], 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 ^[18].