Perplexity brings its local AI agent to Windows, if you have 24GB of VRAM
Portable Computer runs its harness and models on RTX PCs, adds local MCP and scheduled jobs, and can hand selected work to cloud models.
By Ryan Merket · Published
Primary source: Perplexity on X
Why it matters
Perplexity is turning its cloud agent into a hybrid product that can keep sensitive work on user-owned hardware. The 24GB VRAM requirement makes privacy practical for a narrow, well-equipped market first.
Aravind Srinivas (@AravSrinivas), Perplexity's co-founder and CEO, brought the AI search provider's local agent to Windows PCs on Monday, extending Portable Computer beyond its initial Linux release while preserving the option to call cloud models for harder tasks.
https://x.com/perplexity_ai/status/2099514386193027201/video/1
Perplexity said in an X thread on September 14th that Portable Computer is available through its Windows app on PCs equipped with NVIDIA RTX GPUs. On-device inference requires at least 24GB of video memory, a threshold that puts the release outside the reach of most ordinary laptops and desktops.
For machines that clear that bar, Perplexity says the agent can run its models, harness and supporting agents locally. Users can give it access to files and connected applications without sending the task itself to Perplexity's cloud. Portable Computer can still escalate selected work to frontier cloud models when a local model needs stronger reasoning or current information.
That hybrid design is the central bet. Srinivas, a former OpenAI researcher who co-founded Perplexity in 2022 with Denis Yarats, Johnny Ho and Databricks co-founder Andy Konwinski, has spent 2026 pushing Perplexity beyond search responses and into software that executes work. Portable Computer moves the execution boundary onto hardware controlled by the user, while leaving Perplexity's cloud available as an optional source of additional compute and web access.
The Windows release adds unattended local work
The Windows version also introduces local Model Context Protocol support, allowing users to connect Portable Computer to their own tools and application integrations. Perplexity has added scheduled tasks as well, so recurring jobs can run on the PC while its owner is away.
Scheduling makes the product closer to an always-on worker than a desktop chatbot. A user could leave Portable Computer monitoring local documents, processing a recurring export or preparing work through connected services. It also raises the stakes for permissions: an unattended agent with access to local files and external applications carries broader authority than one answering a prompt in an open window.
Perplexity says code and tool execution operate in isolated sandbox environments. Portable Computer asks for permission before sending device content to a cloud service, according to Perplexity's August 25th launch post. Local execution does not consume Perplexity credits, though tasks that invoke web search, connected applications or cloud models still cross the device boundary.
The original Linux release ran on NVIDIA's DGX Spark and supported Qwen 3.8 27B alongside PPLX 27B, Perplexity's post-trained version of the Qwen model. Perplexity said the orchestrator, planner, tool router, scheduler, durable task queue and local search index all ran on the device. Connectors included Google Drive, Gmail, Slack and GitHub.
NVIDIA outlined the Windows expansion on September 3rd, describing Portable Computer as one of several agent applications being adapted for simpler local setup on RTX hardware. The chipmaker said the setup uses llama.cpp with NVIDIA inference optimizations and identified the same 24GB memory floor. The Windows release arrived 11 days later, turning NVIDIA's "coming soon" commitment into a shipping product.
Local models still trail their cloud counterparts
Perplexity's own research provides a useful limit on the privacy and cost pitch. In a 53-task internal knowledge-work benchmark, Perplexity reported that its Computer harness scored 82.6% with Qwen 3.8 27B and 85.4% with PPLX 27B. Those figures came from Perplexity's benchmark and have not been independently validated.
The same research report found a clearer gap on difficult coding jobs. A local Qwen model scored 59.6% on Terminal Bench 2.1 in Perplexity's testing. Adding advice from a cloud model raised the score to 73%, while the cloud model running alone reached 82.4%.
That performance gap explains why Portable Computer retains a cloud escape hatch. Perplexity can keep routine file handling and agent orchestration on a user's RTX machine, then sell access to stronger remote models when local inference falls short. NVIDIA benefits from the same arrangement: every agent that requires 24GB of video memory gives customers another reason to buy higher-memory GPUs and workstations.
Windows support substantially expands the potential installation base from the Linux-first launch, though the hardware requirement keeps Portable Computer aimed at developers, AI enthusiasts and organizations already paying for capable NVIDIA systems. Perplexity has shipped the software path to local agents on Windows. Broad adoption now depends on how quickly suitable hardware moves from specialist workstations into ordinary PCs.