--- title: Perplexity's Computer Agent Goes Local, But You'll Need... description: Perplexity launched Portable Computer, a local-only version of its agentic Computer assistant built with Nvidia, requiring an RTX GPU with at least 24GB VRAM... canonical: https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Perplexity's Computer Agent Goes Local, But You'll Need Serious Hardware | daily.dev og:description: Perplexity launched Portable Computer, a local-only version of its agentic Computer assistant built with Nvidia, requiring an RTX GPU with at least 24GB VRAM... og:url: https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg og:image: https://api.daily.dev/og/posts/FlzIZ79Bg.png og:image:alt: Perplexity's Computer Agent Goes Local, But You'll Need Serious Hardware og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Perplexity's Computer Agent Goes Local, But You'll Need Serious Hardware **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary Perplexity launched Portable Computer, a local-only version of its agentic Computer assistant built with Nvidia, requiring an RTX GPU with at least 24GB VRAM or an Nvidia DGX Spark. The whole runtime, orchestrator LLM, subagent LLM, and agent harness runs on-device with no cloud dependency by default. Because the local model (Qwen3.8-27B) struggles past 100,000 tokens despite a claimed 260K context window, Perplexity built a deterministic orchestrator with an OS-level sandbox that escalates to cloud models only with user approval when needed. It reportedly beats Pi and Hermes on internal benchmarks. It's live now for Ubuntu on RTX/DGX hardware for Pro, Max, and Enterprise subscribers; Windows support arrives in September, and Mac support has no timeline. ## Content Perplexity launched Portable Computer this week, a fully local version of its Computer agent that runs on Nvidia RTX GPUs (24GB+ VRAM) or the DGX Spark desktop, starting at $4,700. The entire runtime - orchestrator, subagent LLM, and agent harness - runs on-device with no cloud dependency by default. The core architectural decision worth paying attention to: rather than making the model smarter, Perplexity made the orchestrator dumber on purpose. A deterministic runtime assembles context, enforces policy, and executes approved tool calls inside an OS-level sandbox. The model only proposes actions. The sandbox is the actual security boundary - tool calls are disabled if unavailable, not just blocked by policy. Building this required reworking the harness to work with smaller local models. They landed on Qwen3.8-27B, which runs well locally but struggles past 100,000 tokens despite a 260K context window. When the local model can't finish a task, the system can escalate to a cloud model - but only with explicit user approval. ## Benchmark results Perplexity tested the same Qwen3.8-27B base model across three agent stacks on a 53-task internal benchmark: - **Computer:** 82.6% - **Pi:** 77.6% - **Hermes:** 74.0% The gap widens on ParseBench-100, a document understanding benchmark covering tables, charts, layout, and formatting: - **Computer:** 65.1% - **Pi:** 34.6% - **Hermes:** 13.9% The ParseBench numbers are notable because they isolate harness quality from model quality - same model, very different results. Document parsing is often the first step in knowledge work, and apparently the orchestrator design matters a lot there. Local reasoning still has a ceiling. On Terminal Bench 2.1, Computer scored 59.6% locally, rising to 73.0% when allowed to consult Claude Opus 5 remotely. Opus 5 alone scores 82.4%, so the hybrid approach closes most but not all of the gap. ## Availability Portable Computer is available now for Ubuntu on RTX and DGX hardware for Perplexity Pro, Max, and Enterprise subscribers. Windows support is coming in September. No Mac support is planned. One fair criticism: calling this "local-first" or "on-device" when the reference hardware is a $5,000 dedicated AI workstation stretches those terms. For most people, this is a cloud-to-enterprise-hardware shift, not a laptop feature. That said, the architectural separation between probabilistic reasoning and deterministic control is a real design choice with real benchmark support - not just marketing. ## Questions this post answers ### What hardware do I need to run Perplexity's Portable Computer agent locally? An Nvidia RTX GPU with at least 24GB of VRAM, or an Nvidia DGX Spark desktop, which starts at $4,700. The entire runtime, including the orchestrator, subagent LLM, and agent harness, runs on-device with no cloud dependency by default, though users can opt to escalate tasks to a cloud model with explicit approval. _daily.dev tracks releases like this for teams weighing local versus cloud AI agent deployments._ ### How does Perplexity's local Computer agent compare to other agent stacks using the same base model? Using the same Qwen3.8-27B model, Computer scored 82.6% on a 53-task internal benchmark versus 77.6% for Pi and 74.0% for Hermes. The gap widened sharply on ParseBench-100, a document understanding benchmark, where Computer hit 65.1% against 34.6% for Pi and 13.9% for Hermes, isolating harness design as the differentiator rather than model quality. _Comparing agent harness benchmarks like these helps developers choosing between orchestration stacks, a task daily.dev surfaces well._ ### How much does local AI agent performance lag behind cloud models like Claude Opus 5? On Terminal Bench 2.1, Perplexity's Computer agent scored 59.6% running fully locally, rising to 73.0% when allowed to escalate to Claude Opus 5 remotely. Claude Opus 5 alone scores 82.4% on the same benchmark, meaning the hybrid local-to-cloud approach closes most but not all of the performance gap. _Developers weighing local versus cloud AI tradeoffs can follow benchmark comparisons like this one on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 1 comments across x (as of 2026-09-13). **TL;DR:** There's barely any discussion to go on; the one substantive reply raises skepticism about the term 'perplexity' as a model metric rather than commenting on the product itself. **Sentiment:** 0% positive · 40% mixed · 60% skeptical **The pushback** - One reply suggests high perplexity scores can signal a model struggling to generalize, questioning the metric's implications. **By community** - x (skeptical): The lone substantive reply casts doubt on perplexity as a metric, seeing it as a potential red flag for model generalization. **Highlights** > @jerryjliu0 @perplexity_ai Perplexity can be a red flag in language models, especially with a fixed vocabulary like LLaMA. Its often a sign the model is struggling to generalize. > — [Michael\_WCD on x](https://x.com/Michael_WCD/status/2092646583158309176) **Source threads** - [x](https://x.com/jerryjliu0/status/2092640502726406310) · 0 points · 0 comments - [x](https://x.com/jerryjliu0/status/2092642169614200989) · 0 points · 1 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#nvidia](https://daily.dev/tags/nvidia), [#local-ai](https://daily.dev/tags/local-ai), [#qwen](https://daily.dev/tags/qwen), [#perplexity](https://daily.dev/tags/perplexity) [View this post on daily.dev](https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Perplexity's Computer Agent Goes Local, But You'll Need Serious Hardware","url":"https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg"},"datePublished":"2026-08-25T15:41:07.208Z","dateModified":"2026-09-13T19:30:02.044Z","description":"Perplexity launched Portable Computer, a local-only version of its agentic Computer assistant built with Nvidia, requiring an RTX GPU with at least 24GB VRAM...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/118e3ed201785b68ec68d5be4e51fdb5?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/118e3ed201785b68ec68d5be4e51fdb5?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,nvidia,local-ai,qwen,perplexity","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Perplexity's Computer Agent Goes Local, But You'll Need Serious Hardware"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/perplexity-s-computer-agent-goes-local-but-you-ll-need-serious-hardware-flziz79bg#faq","mainEntity":[{"@type":"Question","name":"What hardware do I need to run Perplexity's Portable Computer agent locally?","acceptedAnswer":{"@type":"Answer","text":"An Nvidia RTX GPU with at least 24GB of VRAM, or an Nvidia DGX Spark desktop, which starts at $4,700. The entire runtime, including the orchestrator, subagent LLM, and agent harness, runs on-device with no cloud dependency by default, though users can opt to escalate tasks to a cloud model with explicit approval. daily.dev tracks releases like this for teams weighing local versus cloud AI agent deployments."}},{"@type":"Question","name":"How does Perplexity's local Computer agent compare to other agent stacks using the same base model?","acceptedAnswer":{"@type":"Answer","text":"Using the same Qwen3.8-27B model, Computer scored 82.6% on a 53-task internal benchmark versus 77.6% for Pi and 74.0% for Hermes. The gap widened sharply on ParseBench-100, a document understanding benchmark, where Computer hit 65.1% against 34.6% for Pi and 13.9% for Hermes, isolating harness design as the differentiator rather than model quality. Comparing agent harness benchmarks like these helps developers choosing between orchestration stacks, a task daily.dev surfaces well."}},{"@type":"Question","name":"How much does local AI agent performance lag behind cloud models like Claude Opus 5?","acceptedAnswer":{"@type":"Answer","text":"On Terminal Bench 2.1, Perplexity's Computer agent scored 59.6% running fully locally, rising to 73.0% when allowed to escalate to Claude Opus 5 remotely. Claude Opus 5 alone scores 82.4% on the same benchmark, meaning the hybrid local-to-cloud approach closes most but not all of the performance gap. Developers weighing local versus cloud AI tradeoffs can follow benchmark comparisons like this one on daily.dev."}}]} ```