发生了什么
SiliconANGLE reports that AI launched Portable Computer, an on-device version of its cloud-based Perplexity Computer agent. The new software is initially designed to run on Nvidia DGX Spark desktop computers, using local Nvidia hardware to perform multi-step tasks such as online research, data science and coding. The report says the system can divide lengthy tasks among subagents, verify responses and, when necessary, request assistance from a cloud-based neural network.
SiliconANGLE reports that AI launched Portable Computer as an on-device version of Perplexity Computer, a cloud-based introduced in February. According to the report, both products are intended to carry out multi-step work, including developing web applications. When a task becomes especially time-consuming, Perplexity Computer can reportedly split it into smaller pieces and assign them to separate subagents. SiliconANGLE says Portable Computer brings similar agent capabilities to Nvidia’s DGX Spark desktop computer.
The report says the initial hardware target is the DGX Spark, which SiliconANGLE describes as a desktop system with a graphics card based on Nvidia’s Blackwell architecture, a 20-core central processing unit and 128 gigabytes of memory. On launch, the software reportedly uses the open-source Qwen 3.8 27B language model and also supports PPLX 27B, a version optimized for the DGX Spark. SiliconANGLE says Perplexity added multi-token prediction to speed prompt processing, although the article does not provide independent measurements of the resulting performance.
SiliconANGLE reports that Qwen 3.8 27B has a 256,000-token context window but that found the model struggled with prompts exceeding 100,000 tokens in practice. To address that limitation, Portable Computer reportedly includes context compaction, which summarizes long prompts so they fall below the 100,000-token threshold. The product also reportedly includes pre-packaged skills: collections of instructions and other assets intended to improve responses for online research, data science and coding.
The product reportedly combines local processing with selected external connections. SiliconANGLE says Portable Computer can verify the accuracy of responses, ask a cloud-based neural network for help when its on-device model cannot complete a task, and connect to services such as GitHub. The report says the software requests user permission before sending data to external tools and does not give those tools access to on-device files. It also describes a sandbox intended to limit operating-system access and a mechanism that blocks unauthorized network connections. SiliconANGLE reports that plans to extend support to Windows computers with Nvidia RTX graphics cards and add Nvidia’s Nemotron 3.5 Lightning model in a future update.
为什么这很重要
The reported launch places an agent designed for multi-step computer work closer to the user’s own hardware rather than relying entirely on a cloud service. That could matter for users who want local processing, but the source does not establish how much work can be completed locally, how often cloud fallback is needed, or how the system performs against existing alternatives. The reported permission, sandboxing and network controls also make the product’s operating boundaries central to its practical value.
The reported product is significant because it combines an ’s multi-step workflow with local hardware. A conventional chatbot can return an answer, while the product described by SiliconANGLE is intended to break work into subtasks, use specialized skills and connect to software tools. If the reported capabilities work as intended, the user experience could shift from asking for isolated responses toward delegating bounded projects to software that plans and executes several steps.
Local execution may give the product a different privacy and control profile from a purely cloud-based agent. SiliconANGLE reports that Portable Computer runs on Nvidia hardware and does not expose on-device files to connected services without permission. Those controls could reduce unintended sharing in some workflows, but the source does not establish what information is transmitted during cloud fallback, what telemetry the product collects, or whether all integrations enforce the same boundaries. Those omissions limit what can be concluded about privacy from the launch report alone.
The system’s context-compaction feature also highlights a practical constraint of agentic software: a large advertised context window may not translate directly into reliable handling of very long prompts. SiliconANGLE reports that observed difficulty above 100,000 tokens despite the model’s stated 256,000-token window. Summarizing long inputs may help the system operate within that limit, but it can also remove detail or alter emphasis. The source provides no evaluation of how often compaction occurs or whether it changes task outcomes.
The launch could also matter to the Nvidia hardware ecosystem because the initial product is tied to the DGX Spark and is described as supporting a -optimized model. SiliconANGLE says the announcement follows reporting that Nvidia was considering an investment in Perplexity, licensing its technology and hiring key employees. Those reported discussions are not independently confirmed in the source and should not be treated as established facts. The concrete development covered here is the product launch and its reported hardware and software scope.
互动机制：它实际上是如何运作的
以交互方式探索这一发展背后的基础技术。
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
接下来看什么
The main unknowns are pricing, release access, real-world performance and the timing of planned Windows support. SiliconANGLE does not independently confirm ’s technical claims, and the source provides no outside benchmark, user testing or detailed security audit. Future reporting should examine whether the agent’s local model can complete useful tasks reliably, what data leaves the device during cloud fallback, and whether the stated restrictions hold under ordinary use.
Availability and cost are immediate unanswered questions. SiliconANGLE identifies the DGX Spark as the launch platform but does not give a Portable Computer price, a separate software fee, a release timetable beyond the initial launch, or the geographic scope of access. It also says Windows support for Nvidia RTX systems is planned rather than available. Those details will determine whether the product is a broadly usable application or primarily a tool for owners of a specific desktop platform.
Independent testing should focus on task completion rather than model specifications alone. The source reports support for research, data science, coding, subagents and response verification, but it gives no test set, completion rate, latency measurement, error rate or comparison with cloud-based agents. It also does not explain how the system handles failed subtasks, conflicting subagent outputs or inaccurate verification. These are consequential gaps for anyone considering the software for work that affects files, code or external services.
Privacy and security claims require scrutiny in actual deployments. SiliconANGLE reports permission requests, restrictions on access to local files, sandboxing and blocked unauthorized network connections. The article does not identify the sandbox technology, define the permissions in detail, describe how credentials are handled, or report an independent audit. Future evidence should show what data is sent to cloud models and connected services, how users can review or revoke access, and whether a malicious prompt can bypass the stated boundaries.
The planned model and platform expansion will show whether Portable Computer is a single-device product or the start of a wider local-agent strategy. SiliconANGLE reports planned support for Windows systems with Nvidia RTX graphics cards and for Nemotron 3.5 Lightning, described as a 30-billion-parameter mixture-of-experts model optimized for devices with limited processing capacity. The source does not provide a delivery date or performance data for either expansion. SiliconANGLE’s launch account is the available report; the product details and ’s claims have not been independently confirmed here.