Perplexity Launches Portable Computer for Local AI Agents
Perplexity launched Portable Computer on August 25, an on-device version of its Computer agent designed for NVIDIA DGX Spark systems. According to Perplexity, the local stack runs models, orchestration, search, and task infrastructure on the machine, while users can approve individual cloud escalations for web access or frontier-model reasoning. The initial release uses PPLX 27B or Qwen 3.8 27B.
Perplexity launched Portable Computer on August 25, an on-device edition of its Computer agent that runs locally on NVIDIA DGX Spark hardware. According to Perplexity's launch post, the product runs the agent harness, orchestrator, planner, tool router, scheduler, durable task queue, local search index, and local models on the user's machine.
The release is built with NVIDIA and is initially available on DGX Spark, a desktop AI system. Perplexity's blog states that Portable Computer will also become available on NVIDIA RTX GPU PCs. VentureBeat reports that the initial rollout includes DGX Spark and Linux machines equipped with NVIDIA RTX hardware, while ZDNET reports that access is limited to paid Perplexity accounts.
Local models, local state, optional cloud access
Portable Computer initially supports PPLX 27B, a Perplexity post-trained version of Qwen, and Qwen 3.8 27B, according to the company's product and launch pages. Perplexity also lists NVIDIA Nemotron as an upcoming option in its model picker. SiliconANGLE reports that PPLX 27B includes multi-token prediction intended to accelerate prompt processing.
The local agent can analyze files, search documents and code, execute tasks, and retain state across sessions, according to Perplexity. Its product page describes code and tool execution as occurring in isolated environments with controlled access to local files and connected applications.
When a task requires external web access or more capable cloud reasoning, Perplexity states that the orchestrator asks the user before routing that individual step to one of more than 15 cloud models. The result is then returned to the local run. Perplexity states that local work does not consume credits, whereas cloud-routed work can do so.
VentureBeat quoted Nate, Perplexity's vice president of engineering for infrastructure and enterprise, describing the release as bringing the Computer user interface to a fully local app, including the agent harness, inference, and components required to perform work locally.
Hardware and data-bound workflows
The product is aimed at workloads involving files and systems that users may not want to send wholesale to a cloud service, such as codebases, document collections, and business records. Perplexity's launch post states that sensitive data can remain on the device while a user selectively authorizes a cloud step when needed.
SiliconANGLE reports that Portable Computer includes pre-packaged skills for online research, data science, and coding, along with a mechanism intended to verify response accuracy. The same report notes that Qwen 3.8 27B has a 256,000-token context window, but that Perplexity found the model struggled beyond roughly 100,000 tokens in practice and added context compaction to summarize long prompts.
That architecture creates a distinct operational tradeoff from cloud-only agents. Local-first systems can reduce recurring inference charges and limit default data movement, but their usable model capacity, latency, and throughput remain tied to the workstation's GPU memory and compute. In comparable deployments, teams also need explicit policies for the moments when an agent requests permission to use external connectors, web search, or cloud models.
Portable Computer's connectors include Gmail, Outlook, Slack, and GitHub, according to Perplexity's product page. Those integrations make user approval and sandbox boundaries central implementation details rather than secondary settings: a local run can remain local until a task requires an action or information source outside the device.
The launch extends Perplexity Computer, the company's cloud-based agentic platform introduced in February. VentureBeat describes the original platform as orchestrating models, files, tools, and web access for multi-step knowledge-work tasks. Portable Computer packages much of that agent stack for local execution, while retaining an approval-gated path to cloud models for tasks the local system cannot complete.
Key Points
- 1Portable Computer moves Perplexity's agent harness and local models onto DGX Spark, reducing dependence on per-token cloud inference for local work.
- 2Users can authorize individual cloud steps for web access or frontier reasoning, creating a hybrid execution model rather than a fully disconnected agent.
- 3Comparable local-first deployments make hardware capacity, sandboxing, connector permissions, and long-context handling important engineering considerations for practitioners.
Scoring Rationale
The release is a notable agent-platform update because it packages local inference, orchestration, tool routing, and approval-gated cloud escalation into one product. It is most relevant to practitioners evaluating private-data workflows and local agent infrastructure, although initial hardware and account requirements limit near-term reach.
Sources
Primary source and supporting public references used for this report.
View 4 more sources
- Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costsventurebeat.com
- Portable Computer is Perplexity's new local AI agent - why it's a game changerzdnet.com
- Perplexity Launches Local AI Model That Will Run on Your GPU Instead of the Cloudgizmodo.com
- Perplexity AI launches Portable Computer on-device AI agentsiliconangle.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
