Perplexity is releasing a local version of its AI agent, Computer. Portable Computer can analyze files and execute workflows without sending data to the cloud. For tasks that require more computing power or up-to-date information, the system can still utilize cloud services.
The new version will initially run on the Nvidia DGX Spark and will later become available for PCs with Nvidia RTX GPUs. Perplexity is thus opting for a hybrid approach in which local processing is the default, rather than the cloud.
Sensitive data remains local
Portable Computer is designed for tasks where data cannot simply be sent to an external AI service. The system can search local files and code, combine information from various documents, and perform actions on the computer. Components such as the planner, tool router, scheduler, and search index also run locally.
Users can choose between Qwen 3.8 27B and PPLX 27B, a variant Perplexity adapted for the DGX Spark. The latter uses multi-token prediction to speed up processing, according to SiliconANGLE.
Perplexity is also attempting to address a practical limitation of Qwen. The model has a context window of 256,000 tokens but would struggle with prompts exceeding approximately 100,000 tokens. Portable Computer therefore uses context compaction: it summarizes long context to limit the amount of information the model must process at once.
The software also includes pre-configured skills for online research, data science, and programming, among other things. It also includes a mechanism to verify the reliability of answers. Tasks executed locally do not consume Perplexity credits.
The cloud remains available
Working entirely locally has limitations, especially when you need up-to-date information or more powerful models. Portable Computer can therefore offload parts of a task to the cloud. This can include web research, browser use, connected applications, or more complex reasoning tasks.
According to Perplexity, more than fifteen cloud models are available for this purpose. The company states that sensitive data can remain local while only the necessary part of a task is processed externally. When information actually needs to be sent from the computer to a cloud service, the user must first give permission.
Integrations with Google Drive, Gmail, Slack, and GitHub are also possible. This allows the agent to combine information from various business applications and then prepare or execute actions locally.
Local speech processing
In addition to files and code, Portable Computer can process speech locally. It uses Nvidia Nemotron 3.5 ASR for this purpose. Both transcription and subsequent file processing can continue on the device.
Code and tools run in isolated sandbox environments. These environments restrict which parts of the operating system models and tools can access. Portable Computer also blocks unauthorized network connections. External services do not have direct access to local files.
Portable Computer is currently available to Pro and Max subscribers who have an Nvidia DGX Spark. This compact AI machine combines a 20-core Arm processor with an Nvidia GPU and 128 GB of shared memory. The first version of the software runs on Linux. Support for Windows and RTX PCs is expected to follow later. Nvidia’s open-source Nemotron 3.5 Lightning model, with 30 billion parameters, is also set to become available.