What happened
SMBtech reports that Perplexity launched Portable Computer, a local-first AI agent application designed for Nvidia’s DGX Spark. The application is described as supporting local reasoning, research and agent workflows, while allowing users to switch between local and cloud-hosted models. SMBtech says the local setup includes a post-trained Qwen 3.8 27B model and that Perplexity is developing a fine-tuned Nemotron 3.5 Lightning variant.
SMBtech reports that Perplexity launched Portable Computer as a local-first AI agent application optimized for Nvidia’s DGX Spark hardware. The report says the application is intended to run reasoning, research and agent workflows on users’ own systems, with files, models and data kept locally for those workflows. The source does not identify a separate Nvidia announcement confirming the launch, and it does not independently verify the product’s release status beyond its reporting.
According to SMBtech, Portable Computer can connect to Google Drive, Gmail, Slack and GitHub. Users can reportedly choose between local and cloud-hosted models: local inference is presented as suitable for routine work, while cloud inference can be used for tasks requiring current information or more complex reasoning. SMBtech says locally run workflows do not count against cloud token limits. The report does not explain what information is sent to cloud services or connected applications, nor does it describe permission, retention or audit settings.
SMBtech says Perplexity provides a one-click local inference setup using a post-trained Qwen 3.8 27B model. The company is also reported to be developing a fine-tuned Nemotron 3.5 Lightning variant intended to provide faster local responses. The article gives no independent benchmark for Portable Computer itself, and it does not establish whether either model is available to every DGX Spark owner or whether additional configuration is required.
The report places the launch within a broader Nvidia push around local open models, developer tools and agent-focused software. It says support for GeForce RTX and RTX PRO GPUs, Windows and DGX Station is planned. Those future-support statements are attributed to SMBtech’s account and should not be treated as confirmed availability or a delivery commitment. The source also discusses other model and software releases, but those are separate items rather than evidence that Portable Computer is already supported across all of those systems.
Read the primary source: smbtech.au ↗
Why it matters
The reported launch could make local AI agents more practical for users handling private files, code and connected work services. Running some workloads locally may reduce reliance on cloud token limits and keep data on a user-controlled system, but the article does not independently confirm the product’s availability, performance, pricing, security controls or supported configurations.
Portable Computer addresses a practical limitation of cloud-based AI agents: sensitive files and work context may need to leave a user’s system before an agent can act on them. If SMBtech’s description is accurate, some tasks could instead run on local hardware, which may appeal to developers, businesses and individuals with privacy or data-residency requirements. Local execution can also reduce dependence on recurring cloud usage limits for suitable workloads.
The reported connection to Google Drive, Gmail, Slack and GitHub makes the product more consequential than an offline chatbot. An agent that can access files, messages, repositories or other work systems may be useful for research and routine tasks, but it also creates a larger permission and security boundary. The source does not say whether the application can take actions in those services, what approvals are required, how credentials are stored, or how access can be revoked.
The local-versus-cloud choice could become an important design pattern for AI applications. Local models may offer greater control over data and predictable access to private context, while cloud models can provide capabilities that depend on fresh information or more computing capacity. That flexibility is potentially useful, but the article provides no comparative measurements of accuracy, latency, cost, energy use or privacy between the two modes.
The practical impact remains uncertain because the source is a secondary report and does not provide product documentation, user testing or an independently verified demonstration. It also does not state pricing, geographic availability, operating requirements, model licensing terms or whether the application is ready for general users. The launch should therefore be understood as a reported product move, not as proof that local autonomous agents are reliable or safe for high-stakes work.
What to watch next
The key questions are whether Portable Computer is broadly available, which DGX Spark configurations it supports, how local and cloud processing are separated, and what data is transmitted to connected services. SMBtech also says support for GeForce RTX and RTX PRO GPUs, Windows and DGX Station is planned, but does not provide a release timetable or independent testing of those versions.
Availability and scope are the immediate unknowns. SMBtech reports a launch for DGX Spark but describes support for GeForce RTX and RTX PRO GPUs, Windows and DGX Station as planned. Follow-up documentation should clarify whether Portable Computer is available now, which hardware is supported, whether it runs on a single device or requires additional services, and whether the planned versions will have the same capabilities.
Data flows deserve close scrutiny. Users will need clear information about which tasks remain local, when cloud inference is invoked, what telemetry is collected, and whether prompts, files or outputs are retained. Connections to Gmail, Google Drive, Slack and GitHub also make permission design important: practical safeguards would include narrowly scoped access, explicit confirmation before external actions, credential isolation and logs that show what the agent accessed or changed. None of those safeguards is confirmed in the source.
The Qwen 3.8 27B and planned Nemotron 3.5 Lightning integrations should be evaluated separately from the application’s marketing description. Useful evidence would include independent tests of coding, research, tool use, long-running tasks and failure recovery on DGX Spark. Performance claims about related Nvidia-supported models elsewhere in the article do not establish how Portable Computer performs in real workflows.
The broader question is whether local hardware can support useful agents without shifting unacceptable maintenance and security burdens to users. Monitoring model updates, software dependencies, connected-service permissions and unexpected actions will matter as much as raw inference speed. SMBtech’s report does not establish how Portable Computer handles errors, prompt injection in retrieved content, malicious files or unsafe agent requests, so those protections remain meaningful unknowns.
