Perplexity launched Portable Computer today, a fully local version of its agentic Computer platform that runs on Nvidia DGX Spark desktops and Linux machines with RTX GPUs carrying at least 24GB of VRAM. All model inference, file access, and task execution stays on the device. Cloud credits bill at zero. The system only escalates to a frontier cloud model with explicit user permission, per step.
The product ships as a single bundled stack: local model weights, agent harness, inference engine, tools, app connectors, and an OS-level security sandbox. That bundling matters because it eliminates the manual assembly that has made local AI painful. Perplexity also published a research paper alongside the launch arguing that small models require co-designed harnesses, not general-purpose frameworks built for frontier scale. Their internal benchmark, Local Knowledge Work Bench, covering 53 tasks across research, financial analysis, and document creation, shows Computer running Qwen 3.8 27B scoring 82.6% versus 77.6% for the Pi harness and 74.0% for Hermes on identical hardware. The paper also documents why Qwen 3.8 27B degrades past 100,000 tokens despite advertising a 260,000-token context window, and how Perplexity built around that constraint using on-demand skill loading and compact CLI tool wrappers instead of MCP servers.
At launch, supported models are Qwen 3.8 27B and Perplexity's own post-trained PPLX 27B, which scores 85.4% on the same benchmark. Nvidia's Nemotron 3.5 Lightning follows soon. Windows support arrives in September. The research paper's benchmark suite is slated for open-source release. Read the paper for the BrowseComp and multimodal results, where the performance gaps between harnesses widen considerably and the token efficiency numbers become harder to dismiss.
[READ ORIGINAL →]