代理系統
Perplexity Moves Its Full Agent Stack to DGX Spark, Sending Only Difficult Steps to the Cloud with User Approval
Portable Computer runs a 27B model, planner, tool routing, persistent task queue, and search index locally, while permission prompts control data transfers. The initial release supports only Linux and DGX Spark, and its official results come from an unpublished internal workload that should not be treated as independent validation.
Perplexity has launched Portable Computer, moving its previously cloud-dependent Computer agent stack to NVIDIA DGX Spark. Qwen 3.8 27B or the post-trained PPLX 27B, along with the planner, tool router, scheduler, persistent task queue, and local search index, all run on the device; code and tools operate in isolated sandboxes. DGX Spark’s GB10 platform provides a 20-core Arm CPU, a GPU, and 128GB of unified memory—enough to accommodate the model and long-running agent state.
The system is “local-first” rather than fully offline. The agent can read files, search code repositories, and generate documents locally. Only when it needs real-time internet data, a browser, connectors, or stronger reasoning does it recommend handing a specific step to one of more than a dozen cloud models. If device content must be sent to services such as Gmail, Slack, or GitHub, the interface requests authorization first. Local speech transcription is also available through Nemotron 3.5 ASR, avoiding the need to upload recordings beforehand.
The key technical feature of this layered approach is that tool-execution authority remains local, while cloud models primarily provide recommendations instead of receiving direct access to the entire file system. However, users must still inspect approval screens to see exactly what context they include. “Ask first” does not mean data minimization has undergone an external audit, and connectors remain subject to third-party permission models.
Perplexity says PPLX 27B scored 85.4% on the 53-task Local Knowledge Work Bench, compared with 82.6% for the base Qwen model using the same harness. However, the benchmark has not been published, and the search evaluations used different backends, so the gap cannot currently be attributed entirely to either the model or the agent architecture. The initial release is available only to Pro and Max subscribers on DGX Spark running Linux; support for RTX, Windows, and Nemotron models is still forthcoming. Engineering teams should next look for public benchmarks, data-egress audit logs, offline failure modes, and evidence that local hardware depreciation and electricity costs are genuinely lower than cloud API costs.