On August 25, 2026, Perplexity announced the launch of Portable Computer, a local edition of its agent platform, Computer, designed to run AI tasks on hardware owned by the user instead of executing them entirely in the cloud. Availability begins on Nvidia DGX Spark computers and Linux devices equipped with Nvidia RTX graphics processing units, while the company said Windows support will follow in September.
The product works with Pro, Max, Enterprise Pro, and Enterprise Max subscriptions. It requires an RTX card with at least 24GB of video memory, roughly equivalent to a GeForce RTX 3090 or newer. This requirement means that running the platform locally is not available on most consumer computers, although Perplexity presents it as a way to make running advanced agents simpler than assembling their components manually.
A Local Package Instead of Multiple Configurations
Portable Computer combines local models, an inference engine, a task-orchestration agent, tools, application connectors, and a secure sandbox environment in a single application. The system can review folders of documents, analyze data, prepare reports, and access services such as Google Drive, Gmail, and GitHub, then send results to Slack through the available connectors.
In a demonstration presented by the company, an agent running the Qwen 3.8 27B model reviewed a set of 1099 forms and investment documents on a DGX Spark device and identified cases of unnecessary fee payments. Because the execution took place on the device, the cloud-usage credit counter remained at zero. Perplexity says every task starts locally, and no step moves to a more powerful cloud model unless permission is requested.
Local Models and a Purpose-Built Design for Agents
At launch, the available models are Qwen 3.8 27B and PPLX 27B, a model that Perplexity later trained to be compatible with its agent architecture, while Nvidia Nemotron 3.5 Lightning will arrive later. The platform uses vLLM to host inference, with an advanced mode that allows users to connect their own inference endpoint.
Perplexity is based on the idea that the local model and the agent architecture should be designed together. According to a research paper published by the company, models such as Qwen 3.8 27B begin to encounter difficulties after roughly 100,000 tokens, despite advertising a context window of 260,000 tokens. The company therefore adopted a short system prompt, a limited number of tools, and “skills” that load when needed, along with less context-intensive command-line tools instead of large MCP connectors.
The platform enforces operating-system-level sandboxing permanently and stops itself if the sandbox environment is unavailable, rather than running tools with the user’s full permissions. When cloud-model assistance is requested, the system scans the context for personally identifiable information and shows the user what will leave the device; the remote model returns text instructions only and does not access local files or tools.
What Changes in Practice?
The most important commercial change is shifting the cost of use from the number of consumed tokens to the capacity of local hardware. An agent that spends hours reviewing documents or repeating an analysis will not accumulate cloud-usage credits during local execution. Keeping data on the device may also benefit organizations handling sensitive financial, legal, or health documents, while responsibility for securing and managing the device remains with the user or organization.
Perplexity presents a hybrid model rather than complete separation from the cloud. In the Terminal Bench 2.1 test, the local Qwen scored 59.6% at a marginal cost close to zero, and the score rose to 73.0% when assisted by a Claude Opus 5 consultant at an estimated cost of $0.415 per task, compared with 82.4% for the cloud model alone at a cost of $0.65.
The company says Computer achieved 82.6% in its internal Local Knowledge Work Bench test, consisting of 53 tasks, compared with 77.6% for the Pi architecture and 74.0% for Hermes with the same model, while PPLX 27B raised the score to 85.4%. However, these results come from Perplexity’s own evaluations and require independent verification. The company also acknowledges that local models remain below advanced models in difficult reasoning and that cloud escalation narrows the gap without closing it completely.
Limitations and the Significance of the Partnership
The launch connects Perplexity’s goal of offering more private agents that are less dependent on per-token billing with Nvidia’s goal of making devices such as DGX Spark practical tools for running local AI. The two companies had previously announced, in June 2025, a collaboration related to sovereign AI models for European publishers and telecommunications companies.
Nevertheless, at launch Portable Computer remains limited to Linux and Nvidia hardware with a high minimum video-memory requirement, and it includes no announced path for supporting Apple silicon. The product therefore does not represent a general replacement for cloud services or tools such as Ollama, but rather an attempt to package the entire agent and inference layer into a single experience, with a clear trade-off between privacy and cost on one hand and hardware requirements and weaker performance on some tasks on the other.