Perplexity today announced Portable Computer, a local-first version of its agentic Computer platform that runs entirely on hardware users already own. The initial release targets Nvidia DGX Spark desktop supercomputers and Linux machines equipped with Nvidia RTX GPUs, developed in close partnership with Nvidia.
The product is designed so the model, user files and the work itself can remain on the device. Work completed locally consumes no billed tokens, and Perplexity says every task starts on the device by default; the system asks for permission before sending any individual step to a more powerful cloud frontier model.
Nate, Perplexity’s vice president of engineering for infrastructure and enterprise, said at a Monday press briefing: “We’ve basically brought the exact same UI to a fully local app.” He added that Portable Computer includes the full agent harness, inference, and everything required to do work locally.
Why this matters for Nvidia
For Nvidia — which has spent recent years promoting trillion-dollar AI data centers — the announcement signals a strategic shift: local AI has moved from hobbyist experiments to practical tools, and Nvidia wants to sell the hardware that runs them.
Nader, Nvidia’s director of developer technology, argued that local AI reached an inflection point as many new open-source models became genuinely useful beyond tiny, heavily quantized hobbyist variants.
How Portable Computer packages a full local AI stack
Perplexity Computer, the company’s agentic platform for knowledge work, orchestrates models, files, tools, and web access to complete multi-step tasks like document review, data analysis, report generation and pushing results into business systems. Portable Computer replicates that experience locally by bundling local models, an agent harness, an inference engine, tools, app connectors and an OS-level security sandbox into a single system.
Perplexity says the bundling is the point: most local AI stacks today require users to assemble components separately — downloading model weights, standing up an inference server, wiring together tools and tuning performance. Nate said Portable Computer focuses on making the experience straightforward so users can get up and running quickly.
Demos and hybrid capabilities
In one demo, the system acted as a retail investor reviewing a folder of 1099s and investment documents — sensitive financial materials many users would avoid uploading to a cloud service. Running a 27-billion-parameter Qwen model at full GPU utilization on a DGX Spark, the agent reviewed each document and flagged instances of unnecessary fees. Nate noted that the cloud-credit counter in the UI remained at zero because all processing occurred on the device.
A second demo illustrated the hybrid model: the agent analyzed a CSV of user funnel data locally, then pushed the finished analysis to a Slack channel using Perplexity’s connector ecosystem, demonstrating that local-first does not mean disconnected.
The system connects to Google Drive, Gmail and GitHub and can escalate to a cloud frontier model when local models reach their limits. At launch users can set up Qwen 3.8 27B or PPLX 27B (a Perplexity post-trained version), with Nvidia’s Nemotron 3.5 Lightning coming soon.
Availability and hardware requirements
Portable Computer is available today for Pro, Max, Enterprise Pro, and Enterprise Max subscribers on Linux; Windows support is scheduled for September. Any RTX GPU with at least 24 GB of VRAM — roughly a GeForce RTX 3090 or newer — meets the minimum requirement, which Perplexity describes as the floor for delivering a good experience while keeping the product broadly available.
Why co-designing model and harness matters
Perplexity published a research paper arguing that effective local agents require co-design of the model and the agent harness — the scaffolding of prompts, tools and orchestration logic around a model. General-purpose harnesses assume frontier models that can absorb enormous contexts and manage sprawling tool surfaces; small local models often buckle under those demands.
Perplexity found empirically that models like Qwen 3.8 27B, despite advertising 260,000-token context windows, begin to struggle beyond roughly 100,000 tokens. Consequently, Perplexity built a deliberately minimal harness: a succinct system prompt, a small set of core tools, and capabilities that load and unload as on-demand "skills" instead of residing permanently in context. Popular connectors such as Gmail and GitHub were converted from token-hungry MCP servers into compact command-line tools. The company also added self-verification hooks to monitor task health and enforced always-on OS-level sandboxing; if the sandbox is unavailable, the harness disables itself rather than running tools unprotected.
Benchmarks reported by Perplexity
Perplexity’s reported benchmark results come from its own evaluations. On its internal Local Knowledge Work Bench — a 53-task suite spanning deep research, financial analysis and document creation that Perplexity plans to open-source — Computer running Qwen 3.8 27B on a DGX Spark scored 82.6%, compared with 77.6% for the open-source Pi harness and 74.0% for Hermes running the same model.
Perplexity’s post-trained PPLX 27B scored 85.4%. Differences widen on harder tasks: on the BrowseComp web-research benchmark, Computer hit 66.7% accuracy versus 50.2% for Pi and 43.9% for Hermes, while using 51% less wall time and 70% fewer tokens than Pi. On multimodal document understanding, Computer scored 65.1% compared with Hermes’ 34.6% and Pi’s 13.9%.
Token economics: why local execution can be cheaper
The strategic case for Portable Computer stems from how AI workloads have shifted. Chat interactions are bursty and short, while agents tend to run continuously and consume many tokens. Nader said the industry is seeing insatiable token demand, and local AI avoids token metering: "you were not metered by the token. You were not paying for the token. So it's really killer for agents."
An agent that runs for hours reviewing documents and iterating on analysis would accumulate substantial API bills in the cloud; on a device the user already owns, marginal token cost approaches zero. Perplexity argues that at enterprise scale token expenditure and data movement become harder to govern, and local-first execution addresses both spend and privacy.
A hybrid example: on Terminal Bench 2.1, a challenging coding benchmark, the fully local Qwen model scored 59.6% at essentially zero marginal cost. Escalating to a Claude Opus 5 advisor in the cloud raised the score to 73.0% at an estimated $0.415 per task; running the frontier model alone scored 82.4% at $0.65 per task. Escalation recovered roughly three-fifths of the gap to frontier performance at about two-thirds of the cost, and the harness requires user consent before sending context out of the device. The remote model returns text guidance only and never accesses local files or tools.
Relation to Ollama and DIY local stacks
Asked how Portable Computer compares to local inference tools like Ollama, Nate drew a distinction: these tools address different layers. Portable Computer’s main effort focused on the agent harness level; underneath it uses vLLM for model inference and offers an advanced mode for users who want to plug in their own inference endpoint. Nate said Perplexity heavily post-trained both Qwen and Nemotron models to achieve the best results and that their focus is on honing the whole stack from model inference to harness.
Nader added that getting inference running quickly on a Spark is achievable (including with tools like Ollama), but as tasks become more agentic you need more performance, different models and different harnesses — "the deeper you go, the deeper it gets." An attendee who struggled to set up his own DGX Spark said the product feels like the unlock needed to let people understand what agentic means and that the right UX is essential.
Nvidia also emphasized scalability: connecting two Sparks over shared memory can run frontier-class open models; four Sparks can run GLM 5.2 or Nemotron Ultra, and Nader said he has seen eight Sparks connected.
Partnership dynamics and open questions
The launch extends an Nvidia–Perplexity partnership that has been growing for more than a year. In June 2025 the companies announced cooperation to bring sovereign AI models to European publishers and telecoms. The sovereign-AI argument — that data belongs to local populations and cultures — echoes Portable Computer’s pitch at the individual-desktop scale: intelligence you control, running on hardware you own.
There is mutual commercial logic. Perplexity, which has raised capital amid rising valuations while facing legal pressure from publishers over content practices (including a lawsuit filed by The New York Times in December 2025 and a prior dispute with Forbes), gains a product whose economics do not rely on metering every token and a differentiated entry into privacy-sensitive enterprises. Nvidia gains a clear commercial use case for DGX Spark, a device attendees said is easier to buy than to operate; when asked whether Spark might ship with Portable Computer and a Nemotron model preinstalled, Nader declined to commit but did not rule it out.
Remaining issues include dependence on Perplexity’s internal benchmarks, the fact that compact models still trail frontier models on hard reasoning tasks (advisor escalation narrows but does not fully close the gap), the Linux-only launch, the 24 GB VRAM minimum that excludes many consumer PCs, and the absence of Apple Silicon from the roadmap. Nate said the company is currently very focused on Nvidia hardware.
Conclusion
Perplexity frames Portable Computer as part of a broader shift in which increasingly capable agents move from remote infrastructure to individual, local devices. Both Perplexity and Nvidia are betting that continued improvements in chips and open models will expand what a desk-side box can do. In the demos the most telling detail was not a benchmark score but the cloud-credit counter sitting motionless at zero while the agent processed a folder of tax documents. For an industry that has measured ambitions in gigawatts and tokens per dollar, Portable Computer proposes a different meter — one that never runs.
