Analysis
Perplexity and Nvidia jointly launched Portable Computer, a fully local version of Perplexity's agentic Computer platform that runs entirely on a user's own hardware rather than cloud infrastructure, VentureBeat reported. Pulse previously covered Perplexity's broader product and funding trajectory, and this launch extends that push into local, on-device inference. The product bundles models, an inference engine, tool connectors, security sandboxing and the agent-orchestration layer into a single application, extending a partnership between the two companies that began in June 2025 around sovereign AI initiatives.
The hardware requirement is real but not exotic: an RTX GPU with at least 24GB of VRAM -- roughly a GeForce RTX 3090 or newer -- or Nvidia's DGX Spark desktop supercomputer, with multiple Sparks able to link over shared memory for larger models. The product launched Linux-only this month, with Windows support arriving in September; Apple silicon support is not on the current roadmap, a meaningful gap given how much AI-development and prosumer usage happens on Mac hardware.
- Perplexity -- providing the agent-orchestration software layer (Computer) and its custom PPLX 27B model variant
- Nvidia -- providing the local-hardware target (DGX Spark, RTX GPUs) and its Nemotron 3.5 Lightning model, coming soon to the platform
- Ollama -- the closest existing competitor on local inference, though VentureBeat's reporting frames the distinction as Ollama solving the inference layer while Portable Computer targets the full agent-harness layer on top of it
- Qwen 3.8 27B -- one of the supported open models, alongside Perplexity's own PPLX variant
“Perplexity's benchmarks claiming advantages over open-source harnesses like Pi and Hermes also remain internal and haven't been independently validated by third parties.”
"Zero token costs" specifically means inference running locally doesn't accrue the per-call API billing that makes extended agentic tasks -- document review, verification loops, iterative multi-step analysis -- expensive to run against cloud models. Perplexity's own testing reportedly found that models like Qwen struggle with reliable performance past roughly 100,000 tokens of context despite being marketed with 260,000-token windows, which is part of the rationale for a purpose-built local agent harness rather than a general-purpose framework layered on top of raw inference.
The honest limitation is that local execution shifts cost from per-token API billing to upfront hardware cost -- a 24GB-VRAM GPU or DGX Spark is not a trivial purchase for an individual user, and "zero token costs" only looks attractive relative to sustained heavy cloud-API usage, not against a user's total cost of ownership including the hardware itself. Perplexity's benchmarks claiming advantages over open-source harnesses like Pi and Hermes also remain internal and haven't been independently validated by third parties.
Users who retain the option to escalate specific complex tasks to Perplexity's frontier cloud models, with explicit permission and cost transparency, get a hybrid model that hedges the tradeoff -- local-first for routine tasks, cloud on demand for anything the local models can't handle reliably, which is closer to how the product will likely get used in practice than either pure-local or pure-cloud framing suggests.
