{"slug": "perplexity-launches-its-computer-agent-locally-on-nvidia-s-4699-desktop", "title": "Perplexity launches its Computer agent locally on Nvidia's $4,699 desktop", "summary": "Perplexity launched Portable Computer on August 25, moving its Computer agent to run fully local on Nvidia's DGX Spark desktop, which costs $4,699, with zero per-token costs for local jobs. The initial release supports Qwen 3.8 27B and PPLX 27B models, with Nvidia's Nemotron 3.5 Lightning coming later, and requires a Perplexity subscription (Pro at $20/month or Max at $200/month). This shift lets users run heavy tasks without consuming cloud credits, which are valued at 100 credits per $1 and typically cost 350–950 credits per complex task.", "body_md": "# Perplexity launches its Computer agent locally on Nvidia's $4,699 desktop\n\n**Portable Computer keeps models and files on a DGX Spark, while cloud escalation still requires approval and credits.**\n\nBy [Ryan Merket](/author/ryan-merket)\n· Published\n\nPrimary source: [VentureBeat](https://venturebeat.com/infrastructure/perplexity-partners-with-nvidia-to-launch-portable-computer-a-fully-local-ai-agent-with-zero-token-costs)\n\n## Why it matters\n\nPerplexity is shifting agent economics from metered cloud inference to customer-owned Nvidia hardware, giving Nvidia a practical application for DGX Spark and Perplexity a way to support heavier usage without carrying every token.\n\n[Perplexity launched Portable Computer](https://www.perplexity.ai/sv/hub/blog/introducing-portable-computer-for-local-first-ai) on August 25th, moving the models, orchestration software and security sandbox behind its Computer agent onto Nvidia hardware. Local jobs carry no per-token charge, although access still requires a Perplexity subscription and a machine capable of running the models.\n\nThe initial release runs on the [Nvidia DGX Spark](https://www.nvidia.com/en-us/products/workstations/dgx-spark/), a desktop system with a Grace Blackwell GB10 chip and 128 GB of unified memory. Nvidia currently lists the machine at [$4,699](https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/), putting a considerable hardware bill behind Perplexity's pitch for unmetered inference. Perplexity Pro costs $20 a month, while Max costs $200 a month.\n\nThat trade moves agent spending from recurring cloud credits toward hardware that customers own. Perplexity's [credit documentation](https://www.perplexity.ai/help-center/en/articles/13838041-how-credits-work-on-perplexity) values 100 Computer credits at $1 and says complex tasks typically consume 350 to 950 credits, with larger projects costing considerably more. Portable Computer lets users run repeated document analysis, code work and other long jobs without drawing down those credits when every step remains local.\n\n[Nvidia's account of the launch](https://blogs.nvidia.com/blog/local-ai-open-source-models-agents-nemotron/) says Portable Computer is available on DGX Spark, with support for GeForce RTX and RTX Pro machines coming later. Perplexity's own launch post also describes RTX PC support as forthcoming. That is narrower than [VentureBeat's report](https://venturebeat.com/infrastructure/perplexity-partners-with-nvidia-to-launch-portable-computer-a-fully-local-ai-agent-with-zero-token-costs), which said Linux computers carrying Nvidia GPUs with at least 24 GB of VRAM were supported at launch.\n\n### Perplexity rebuilt the harness for smaller models\n\nPortable Computer packages the model, inference engine, planner, tool router, scheduler, local search index and sandbox into one application. Users can install it on a DGX Spark instead of assembling a local model server, agent framework and execution environment separately.\n\nNate Kupp, Perplexity's vice president responsible for Computer infrastructure and enterprise engineering, told VentureBeat that the work centered on the agent harness surrounding the model. That layer decides which tools the model can use, what information remains in context and when a task should seek outside help.\n\nAt launch, Perplexity offers Qwen 3.8 27B and PPLX 27B, its post-trained version of the Qwen model. Nvidia's [Nemotron 3.5 Lightning](/models/nvidia/nemotron-3.5-lightning:free) is scheduled to join the model picker later. The system can read local files, execute tools inside isolated environments and connect to services including Gmail, Slack, Google Drive and GitHub.\n\nA task begins locally. When it needs web access or stronger reasoning, Portable Computer asks the user before sending approved context to one of Perplexity's cloud models. The remote model returns guidance and receives no direct control over local files or tools, according to Perplexity. Those cloud steps can still consume credits.\n\nPerplexity's security design disables tool execution if its operating-system sandbox is unavailable. The company says this prevents the agent from falling back to commands running with the user's full permissions, a dangerous default for software that can edit files and execute shell operations.\n\n### The benchmark numbers remain Perplexity's\n\nIn an accompanying [technical report](https://www.perplexity.ai/hub/blog/a-local-first-agent-for-private-and-cost-effective-knowledge-work), Perplexity said its harness scored 82.6% on an internal set of 53 knowledge-work tasks using Qwen 3.8 27B. The same model scored 77.6% with the open-source Pi harness and 74% with Hermes. PPLX 27B reached 85.4%.\n\nThose figures come from Perplexity's own evaluation, including a benchmark that has yet to receive the external scrutiny of established public tests. Perplexity also acknowledged a hard ceiling for local models: even with a tailored harness, compact models continue to trail frontier systems on demanding reasoning tasks.\n\nThe company's BrowseComp results offer a more comparable measure. Perplexity reported 66.7% accuracy for Computer, against 50.2% for Pi and 43.9% for Hermes, although Computer used Perplexity's search infrastructure while the other harnesses used Brave. That difference makes the result a test of the full stack rather than the harness alone.\n\nFor Perplexity, Portable Computer reduces the cloud inference bill attached to customers who run agents continuously while keeping them inside its subscription and connector products. Nvidia gets a packaged application for DGX Spark, hardware built for local AI that still asks buyers to manage models, inference software and agent frameworks themselves.\n\nPortable Computer turns that box into something closer to an appliance. The token meter may stay at zero, but Nvidia collects the first $4,699.", "url": "https://wpnews.pro/news/perplexity-launches-its-computer-agent-locally-on-nvidia-s-4699-desktop", "canonical_source": "https://runtimewire.com/article/perplexity-portable-computer-nvidia-dgx-spark-local-agent", "published_at": "2026-08-26 18:10:28+00:00", "updated_at": "2026-08-26 18:43:29.676060+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-infrastructure", "ai-agents"], "entities": ["Perplexity", "Nvidia", "DGX Spark", "Qwen 3.8 27B", "PPLX 27B", "Nemotron 3.5 Lightning", "Nate Kupp", "VentureBeat"], "alternates": {"html": "https://wpnews.pro/news/perplexity-launches-its-computer-agent-locally-on-nvidia-s-4699-desktop", "markdown": "https://wpnews.pro/news/perplexity-launches-its-computer-agent-locally-on-nvidia-s-4699-desktop.md", "text": "https://wpnews.pro/news/perplexity-launches-its-computer-agent-locally-on-nvidia-s-4699-desktop.txt", "jsonld": "https://wpnews.pro/news/perplexity-launches-its-computer-agent-locally-on-nvidia-s-4699-desktop.jsonld"}}