本機推論與代理
Perplexity Brings Its Local Agent to Windows RTX, but “Data Stays on the Device” Still Has Cloud Exceptions
Portable Computer can now run models, tools, and sandboxes on Windows RTX PCs with at least 24GB of VRAM, with most local work not counting against cloud quotas. However, the system can still escalate to cloud models with user authorization, while connectors and external actions mean it should not be treated as a fully offline environment.
On September 14, Perplexity added Portable Computer to its Windows app, supporting GeForce RTX and RTX PRO GPUs with at least 24GB of VRAM. This is not merely a chat model wrapped in a desktop app. It combines a local model, agent orchestrator, tools, browser, and proprietary SPACE sandbox into a multistep execution environment. Officially listed use cases include cross-file analysis, scheduled tasks, code execution, and triaging GitHub pull requests.
The system can be configured with a 27B model post-trained for Perplexity Computer and optimized for RTX. NVIDIA’s Windows announcement cites Qwen 3.8 27B as an example; Perplexity previously also listed PPLX 27B, which is based on that model. For deployers, the 24GB threshold means consumer-grade options are largely limited to high-end cards such as the RTX 3090, 4090, and 5090 rather than typical laptop GPUs. The model’s actual quantization format, context length, and tool runtime footprint will still affect the available headroom.
Portable Computer’s key distinction is its hybrid routing. Work runs locally by default, but when a step requires more powerful reasoning, the agent first requests permission and then sends some content to a cloud model. Work completed locally does not consume Computer quota, but that commitment does not cover authorized cloud inference. Connectors for Outlook, OneDrive, Google Drive, Gmail, Slack, and GitHub also establish their own network and permission boundaries. As a result, “sensitive data stays on the device” should be viewed as a workflow design objective rather than a blanket guarantee covering every tool call.
Engineering teams should next examine exactly what context is transmitted during cloud escalation, whether connectors follow least-privilege access, how sandbox isolation is implemented, and whether local and cloud steps can produce auditable records. The companies have not yet published latency, throughput, power-consumption, or task-success-rate figures across different RTX GPUs. Windows availability has been confirmed, but the performance and privacy claims still require independent testing.