Perplexity unveils local AI platform for NVIDIA DGX Spark

Perplexity announced the launch of Portable Computer, a self‑contained AI agent platform that runs entirely on NVIDIA’s DGX Spark hardware. The offering combines the company’s agent harness, planner, tool router and post‑trained language models into a single installable package, allowing every task to start on the local device. When a step requires live web access or advanced reasoning, the system pauses, prompts the user, and forwards that single step to one of more than fifteen cloud models. This hybrid approach preserves data privacy while still tapping external expertise when needed.

The core of Portable Computer is a local inference engine that supports either the Qwen 3.8 27‑billion‑parameter model or Perplexity’s own PPLX 27B variant, which is fine‑tuned for the platform’s orchestrator. Both models run at 3‑bit quantization, with the Qwen model requiring a 17.4 GB download and about 32 GB of RAM, while the upcoming Nemotron 3.5 Lightning model, a 30‑billion‑parameter mixture‑of‑experts, will be delivered at 4‑bit precision and needs roughly 36 GB of RAM. The software package also includes an OS‑enforced sandbox that isolates tool execution, limits filesystem and network access, and disables tool calls if the sandbox cannot be established. Connectors for Gmail, Outlook, Slack and GitHub are built in, and users can bring their own models or inference servers.

Perplexity’s move reflects a broader industry shift toward “local‑first” AI, where enterprises seek to keep data processing on‑premises to reduce latency, control costs, and meet regulatory requirements. By eliminating per‑token charges for locally handled steps, the platform makes long verification loops and large‑scale migrations economically viable on owned hardware. The requirement for a GB10‑class box or an RTX GPU with at least 24 GB of VRAM positions the product for data‑center or high‑end workstation deployments rather than consumer laptops.

In benchmark testing, Portable Computer demonstrated notable gains over open‑source alternatives. On the 53‑task Local Knowledge Work Bench, the Qwen‑based configuration achieved an 82.6 % success rate, surpassing the open‑source Pi harness (77.6 %) and Hermes (74.0 %) on the same model. The PPLX 27B variant lifted performance to 85.4 %. On the BrowseComp suite, the system reached 66.7 % versus 50.2 % for Pi and 43.9 % for Hermes, while using 51 % less wall‑time and 70 % fewer tokens. Visual document understanding on ParseBench‑100 yielded 65.1 % compared with 34.6 % and 13.9 % for the competitors. A hybrid run on Terminal Bench 2.1 showed a fully local score of 59.6 % at effectively zero marginal cost, which rose to 73.0 % when a cloud adviser was invoked at roughly $0.415 per rollout, narrowing the gap to frontier models like Claude Opus 5.

Despite the strong performance numbers, the solution has clear constraints. It currently supports only Linux for Pro‑level subscribers, with Windows slated for a later release and macOS not planned. Only a single DGX Spark node is supported at launch, and clustering capabilities are listed as future work. The hardware prerequisites, GB10 superchip, 128 GB memory, and at least 1 TB storage, make the offering inaccessible to smaller teams. Moreover, the reliance on a sandbox means that any tool that cannot be sandboxed is simply disabled, potentially limiting functionality in edge cases.

Looking ahead, Perplexity plans to open‑source the 53‑task Local Knowledge Work Bench, which could foster community contributions and broader adoption. The roadmap includes support for additional hardware platforms, clustering, and the upcoming Nemotron 3.5 Lightning model. Observers will watch for how the hybrid escalation model balances cost and performance, and whether the zero per‑token pricing model spurs wider migration of enterprise workloads to on‑premise AI.