Perplexity Launches Portable Computer: A Local AI Agent That Runs Entirely on Your GPU

Portable Computer is Perplexity's new local AI agent - why it's a game changer

Follow ZDNET: Add us as a preferred source on Google. ZDNET's key takeaways * Perplexity's new agentic Portable Computer runs AI locally. * Local AI is faster and cheaper than cloud-based models. * But you'll need some beefy hardware for Linux or Windows. Perplexity has unveiled a new version of its Personal Computer agent, released in February, that aims to eliminate some of the drawbacks of using AI in the cloud. Known as Portable Computer, this new option relies on local AI models to handle your agentic requests and assignments without the usual online overhead. Also: I gave Perplexity's agentic AI 5 complex tasks to run on my Mac - and I'll do it again Debuting on Tuesday, Portable Computer offers access to all the standard AI commands, features, and tools through which Perplexity's agent can carry out tasks with little or no intervention on your part. The key part here is that the models, your files, and your work all stay on your machine. You control if and when the AI needs to hop online to achieve its mission. As always, you can still connect to external services like Google Drive, Gmail, Slack, and GitHub to incorporate your cloud-based files and information. Cloud vs. local AI Why use AI locally rather than in the cloud? There are a few benefits. First is performance. Local AI models can run faster and respond to your requests more quickly than their online counterparts. Second is security. Your tasks are handled locally in an isolated sandbox on your machine, reducing the risk of data breaches. Third is cost. Running AI in the cloud chews up credits, tokens, and money. Running it locally doesn't use any billing credits. If any part of a task needs to hop online, only the cloud work eats into your credits and tokens. Also: 5 reasons I use local AI on my desktop - instead of ChatGPT, Gemini, or Claude "Perplexity Portable Computer is a local-first agent," Perplexity said in its blog post. "Sensitive data therefore never leaves the device without permission, and local models carry no inference fee: the system is private and cost-effective by construction." How to try it - and limitations Can anyone try Perplexity's new local AI? Unfortunately, no -- that's where the strict requirements come in. Portable Computer is available only for paid accounts, meaning Pro, Max, Enterprise Pro, and Enterprise Max subscribers. Sorry, free users. For now, Portable Computer is accessible only on Linux systems, but Perplexity promises that it will reach Windows users in September. Sounds good, but there's more. Also: Google's AI-first Googlebook laptop is almost here - can it avoid the Copilot+ PC problem? One downside of local AI is that it typically needs some serious hardware to work. And that's certainly the case with Portable Computer. On the Linux side, Perplexity's local AI requires an Nvidia DGX Spark or other Linux machine with an Nvidia RTX GPU (graphics processing unit). For the operating system, you'll need Nvidia DGX OS or Ubuntu on either an ARM or x64 system. On the Windows side, your PC must be outfitted with an NVIDIA RTX GPU with at least 24GB of VRAM. That's where many desktops may not be up to snuff, including my own. My PC does have an Nvidia RTX GPU, but it comes with only 4GB of RAM. I'm not a PC gamer and don't typically run GPU-intensive tasks, so I didn't equip it with a high-end graphics card. Plus, Nvidia RTX GPUs with 24GB of VRAM run at least $1,500, so you'd need to shell out a few bucks for one of these. Model options As far as the local models themselves, Portable Computer offers a couple of options. You can set up Qwen 3.8 27B, a well-known open-source model. This particular AI model has gotten kudos not just for speed and performance but for its prowess with coding, research, professional work, and more complex agentic tasks. Alternatively, you can tap into a post-trained version of the model that Perplexity has dubbed PPLX 27B. A post-trained AI model is one that has received more refined and specific training beyond its initial instructions. Here, Perplexity post-trains all models on Portable Computer to boost their accuracy and make them more efficient. Also: I tested a Debian Linux loaded with local AI - see if it's right for you Coming soon is Nvidia's Nemotron 3.5 Lightning, which Nvidia said is designed for high-volume and long-running tasks. Naturally, you'll be able to switch among the different models depending on the assignments you give the agentic AI. Whether or not your system and GPU can handle the heavy demands of Portable Computer, this latest development does represent a new option for local AI. Hopefully, the requirements will lessen over time, paving the way for more people to turn to local AI for better performance, tighter security, and lower costs.

I built a fully local agentic AI setup with one open-source app, and it never touches the cloud

Handing an app permission to browse, write files, and execute on your behalf is powerful, but there's a whole category of work that I think you shouldn't give to a cloud-based agent. Basically, I feel like if you wouldn't send it in a DM or post it on social media, then it's probably best left on your machine, such as bank statements, contracts, or health info. Every big agentic cowork tool right now, Claude Code, Antigravity, Codex, etc. ships all that context off to servers out of your control. Eigent is an open-source, local-first alternative to those. I've been testing it for a while, and it does actually accept cloud APIs, but the version that matters for this is Eigent paired with a local model. That's the setup that's going to keep your workflow fully local. Want to stay in the loop with the latest in AI? The XDA AI Insider newsletter drops weekly with deep dives, tool recommendations, and hands-on coverage you won't find anywhere else on the site. Subscribe by modifying your newsletter preferences! Some folders need to be kept off-limits And I finally found a place for them to go I have been running Claude Code, Codex, and Antigravity for months, and I have never once pointed any of them at something on my machine that would bother me if it were published publicly. For example, I use these tools for my UI experiments and vibe-coding projects, as well as my Obsidian vault admin and interactive layer. The vault has meeting notes, pitch ideas, a database of my DVD collection, and so on - so yes, private information, but it wouldn't be the end of the world if those contents got more eyeballs. I deliberately keep my health and finance docs out of it because I want to keep Claude Code in it. I've primarily been using my local LLM with the filesystem access plugin to work on those private documents. It's a dead simple setup and gives you an interactive layer on documents you might not have otherwise understood. Until I found Eigent, a cowork app that lets you run the entire pipeline locally and gives you more workspace features than a local LLM engine. I run this with my tried-and-true Qwen 3.5 9B because it's powerful enough for tool calling but small enough to work within the constraints of my hardware. One app running a whole team of agents This agent splits up and runs the pieces at the same time Close Eigent is an open-source desktop app built on the CAMEL-AI multi-agent framework. It doesn't care what model you pick, but I've been running it locally, hooked up to LM Studio's OpenAI-compatible local server so every call runs on my GPU. Ollama, vLLM, SGLang, and llama.cpp all work the same if you prefer one of those. It ships specialized agents - Developer for terminal and code, Search for browsing, Document for file work, Multi-Modal for images and audio. And then there's the parallel workforce. When I hand it a prompt, a Coordinator agent decomposes it into subtasks and dispatches them across the specialized agents, and these independent branches run concurrently. Each agent keeps its own memory window so context does not bleed between them. The Document agent handles the usual suspects: PDFs, Word, Excel, PowerPoint, HTML, EPUBs, plain text, even images via OCR and audio. So bank statements, contracts, and scanned docs are all fair game. I'm not demonstrating this with my actual private documents, of course, so I set it up with my Obsidian vault. A prompt like "cluster my orphan notes by theme and cross-reference them against my open pitches" spins up multiple agents that all work the vault at the same time. My DVD hunts pull the same trick - Document agent reads my wishlist while the Browser agent hits retailers in an isolated Chromium session, both running in parallel while I watch the task tree fill in. It finishes fast enough, though this will come down to your hardware if you're running the full pipeline locally. Don't expect miracles if you have a smaller GPU like me. The agent clocks in without me It runs itself Eigent has a Triggers system that lets you schedule your work, and there are three types. Schedule Triggers for recurring runs, Webhook Triggers for external systems firing agents in real time, and Agentic Triggers where one agent can wake another based on the outcome of its own task. My DVD hunt Skill was the obvious first candidate. I had been running it manually every couple of weeks. Now it runs every Sunday morning on a Schedule Trigger and drops a dated report into my collection folder before I even start my day. Setting it up was pretty easy - I opened Scheduled in the sidebar, gave it a plain prompt of "run dvd-hunt every Sunday at 8 am and save the report to my collection folder", and it parsed the cadence and wired the Skill up itself. Every scheduled run uses the same local model my manual runs do, so unattended does not mean it reaches for the cloud when I am not looking. Eigent also runs a local Postgres instance in the background and logs every agent message and tool call and file write into it. When I check on the Sunday run, I can pull up exactly what happened - which agent did the reading, which page the browser hit, what ended up in my collection folder, and so on. All of this with no cloud-based dashboard behind it, the entire database is still on my machine. Finally, a local agent stack I can use for the folders I would never hand to a cloud one Local agents used to be a massive compromise, they were way slower and less capable, but at least private. Now, if you're in the same boat as me with 8GB VRAM, you might still be dealing with some slowness, but the "incapable" framing overall is outdated now. Eigent with a local model gave me a workforce that runs in parallel on my GPU, a scheduler that wakes it up without me, and a Postgres instance that keeps the receipts. The folders I wouldn't let a cloud agent near finally have something to work with them. Eigent See at Github Expand Collapse

Perplexity Launches Local AI Model That Will Run on Your GPU Instead of the Cloud

Perplexity, a largely forgotten player in the AI space save for its role as Joe Rogan's personal fact checker, is looking to regain some relevance by asking people to think global, process local. The company announced Tuesday that it is launching Portable Computer, a version of its agentic "Computer" platform that runs entirely locally on your own hardware. Computer was launched by Perplexity back in February to offer users an always-on AI agent that can run autonomously to complete tasks whether you're at the keyboard or not. Computer took the OpenClaw approach, where the agentic platform runs on a dedicated device like a Mac Mini, but the actual AI processing is still cloud-based. Portable Computer walls off the garden completely, bringing the model on device -- meaning it's running entirely on a user's hardware. That requires more processing power, which is why Perplexity's initial release of Portable Computer is limited to Nvidia's DGX Spark desktop supercomputer and Linux machines with Nvidia RTX GPUs. Part of the appeal of the Mac Mini for AI agents like OpenClaw was the bang for your buck on processing power -- a Mac Mini with an M4 chip sets you back about $900. You're going to need to fork over about 5x that price for a DGX Spark. The benefit is that's a one-time fee. You won't have to keep tabs on your tokenmaxxing because there's a credits bill coming at the end of the month, since everything is done locally (except when it isn't, but Portable Computer will ask before pinging the cloud for additional processing power, per Perplexity). The other trade-off, though, is that the model selection will be limited. One of the pitches Perplexity made when it launched Computer was that users would be able to tap into 19 different AI models to tackle whatever prompt you give it. With Portable Computer, it'll launch with Qwen 3.8 27B developed by Alibaba or PPLX 27B, which is the same Qwen model that has been post-trained by Perplexity. Nvidia's Nemotron 3.5 Lightning is expected to be available soon. Despite running locally, Perplexity did make a point to highlight that Portable Computer can still connect to popular apps. Per VentureBeat, the company showed off a demo that included Portable Computer sharing some data analysis directly to a Slack channel on behalf of the user and said that it could also connect to Google Drive, Gmail, and GitHub. The heavy Nvidia connection to Portable Computer is likely not a coincidence. The chip giant is reportedly mulling making an investment in Perplexity that would value the company at $30 billion, so it makes sense that it'd like to see what it's investing in firsthand before taking the plunge.

I turned two mini PCs into a local AI that reads my whole codebase, and now it reviews my code for me

I've been spending some time with two DGX Sparks recently, with DeepSeek V4 Flash, a 284-billion-parameter model, being my go-to model of choice running across both of them. They're connected over a ConnectX-7 cable with 128 GB of unified memory each, and can push anywhere from 30 to 50 tokens per second most of the time. Since then, it's become more than just a box for inference when I need it. In fact, these two mini PCs are now the only AI I trust with actual work, and they've quietly become the most powerful productivity assistant I've ever had, because they run all of it on the desk in front of me. Nothing goes to the cloud, nothing is logged by someone else, and the only recurring bill is electricity. Hermes Agent is what enables most of it; a chatbot can answer questions quickly, and there are already endless amounts of choice that you can pay for every month. This isn't that, and instead, is something that genuinely helps improve my productivity. Nvidia DGX Spark $4699 at NVIDIA Expand Collapse Giving the brain a job was what it needed Your AI is useless if you can't use it for anything Everything that this assistant can do starts with something a lot of other local LLMs can't do: read practically everything at once. The full, million-token context window fits on the two Sparks with gigabytes to spare, as the KV cache runs at NVFP4 which Blackwell dequantizes on the tensor cores. A single context window holds around the full million in roughly 8.5 GiB, making basically all of the difference in both capability and speed. To be fair, 50 tokens per second doesn't sound like a lot, but it's plenty fast for most workloads. Prompt processing runs at 1300 to 1900 tokens a second, so loading a large codebase into context barely takes any time at all. On top of that, a 120,000-token session pushes about 90 GB across the ConnectX-7 link between both machines, which is hidden by the 200 Gb/s transfer speed. Most local LLMs people run have the same issue, where they have a much smaller context than the model actually supports thanks to the VRAM constraints most people are running them in. This one keeps the whole conversation, keeps the whole repository, keeps everything from the entire day, and answers from all of it. I also ran a couple of needle-in-a-haystack tests with a 500,000 token fill of irrelevant text, then asked it to recite a specific detail to see how it tracked long-context sessions, and it recited it back perfectly every time. It helps automate my code testing It splits itself into subagents The reason almost nobody uses a local model as an actual assistant isn't for its speed or how smart it is, but because a lot of genuinely useful work requires accessing real things. And accessing those real things means giving it access to secrets and login details, something most people understandably wouldn't trust to a cloud model. This exact problem is why platforms like Tines' 3B exists, as it allows them to legitimately give an AI agent access to services without giving it the actual keys. Naturally, a local model is still as vulnerable to prompt injection as a cloud model, maybe even more so, which is why I'm still very careful about both how and where my Hermes Agent connects to the internet. It's firewalled off from most of my LAN, and the connections it can make to the wider web are dictated by me rather than by the agent itself. All of my interactions in that context are done through Discord and Telegram, though primarily the latter. I have automated cronjobs set up, and one messages me with tech-related news every morning. However, my favorite is a code reviewing agent that I set up recently, and it genuinely works really well. How it works is really simple; it has an SSH key that's been added to a single Git repository I've maintained for a couple of years, and as it's a multi-architecture full-stack deployment, you can imagine that it's grown substantially in that time. However, that's not all the agent has access to: it has a separate token that allows it to open issues in the repository as well. I message it to evaluate the different modules in the repository, it spins up six different subagents all on the same model, and I get a reply with a list of potential issues 20 to 30 minutes later. I pick from the list which ones to actually file as issues, and then I work on them myself after that. As well, because it's Hermes Agent, it tracks everything that I ask it to improve or change, as its memory is one of the primary features of the harness. Persistent memory lives across sessions in actual files it can read at the start of every conversation, and that includes my preferences, any projects, and even the mistakes that it made the week before that it noted so it doesn't happen again. Even better, when I ask it to do something it was never shown, once it figures it out, it can save it as a skill to load back into context the next time a similar task shows up. All of this has been genuinely really useful, especially the code-review part. Are the responses perfect? No, and that's why I have it highlight the issues to me first rather than automatically filing them to Git. I understand the architecture of my project, and there are plenty of things in it that aren't actually used in production but are a part of the QA process, and it has sometimes treated those demo files as if they were part of prod. But it still means I get a fresh pair of eyes on a project that I've worked on for so long that it can be easy to tunnel vision on, and that kind of external "read" is simply unmatched. I just treat it like a pair programmer, rather than something that does the work for me. Managing the electricity bill is the only downside Idle costs are the highest cost The idle cost of both machines is the difference between a productivity hack and yet another recurring expense, so a Wake-on-LAN setup automated through Home Assistant is next on the cards. With that, I'll be able to spin them up when needed and shut them down again when they're not, as the idle costs of both machines far outweighed the cost of actual inference when they were running. The only downside is that, at 40 to 50 output tokens a second, both can generate a maximum of approximately 4.3 million tokens a day, but that's still a lot of tokens to actually output. Mine doesn't come anywhere near it in most weeks, let alone a month, as the majority of the codebase work that I outlined above is input rather than output. A local assistant is decent for execution, but it should never be self-driven in the way you might be tempted to let it. It can do automated code reviews, research, and other, menial tasks, but you never want it to complete actions on your behalf that actually matter. The more useful it's become over the past while, the more it becomes impossible to miss that it's the case. An appliance should be boring, and honestly, when it works, it absolutely is.

Perplexity and NVIDIA team up to release a local AI agent

Nick Lewis is an editor at How-To Geek. He has been using computers for 20 years --- tinkering with everything from the UI to the Windows registry to device firmware. Before How-To Geek, he used Python and C++ as a freelance programmer. In college, Nick made extensive use of Fortran while pursuing a physics degree. Nick's love of tinkering with computers extends beyond work. He has been running video game servers from home for more than 10 years using Windows, Ubuntu, or Raspberry Pi OS. He also uses Proxmox to self-host a variety of services, including a Jellyfin Media Server, an Airsonic music server, a handful of game servers, NextCloud, and two Windows virtual machines. He enjoys DIY projects, especially if they involve technology. He regularly repairs and repurposes old computers and hardware for whatever new project is at hand. He has designed crossovers for homemade speakers all the way from the basic design to the PCB. Nick enjoys the outdoors. When he isn't working on a computer or DIY project, he is most likely to be found camping, backpacking, or canoeing. Perplexity originally made its name as an AI search engine, but it also has a product called Computer. Computer, which launched earlier this year, is an agentic AI that can complete tasks on your behalf. It is a general-purpose work companion that uses the same interfaces and applications that you do. You describe a job using normal language, then Computer will run subagents as needed to perform research and connect tools. Now, Perplexity is launching Portable Computer. The intent of Portable Computer is similar to Computer, with one fundamental difference: it is local first. At launch, Portable Computer will be able to use Qwen 3.8 27B or Qwen PPLX 27B, a Qwen model that has received additional training. Support for NVIDIA's Nemotron 3.5 Lightning (30B) will be available soon. Portable Computer is designed to run as many things as it can locally, on your own hardware. You can summarize documents, extract information and perform analyses, send files to apps, or any other number of jobs. However, if Portable Computer runs into a task that is too complex for the local model, it will stop and ask you if you'd like to offload to a larger, cloud-based model instead. The local-first design allows you to use AI to handle data that you may not want on the cloud. It also has a practical perk: anything you do locally doesn't consume credits or tokens, and the cost of running a local AI is small once you already own the hardware. Close Why use Perplexity's Portable Computer? Portable Computer will save you a ton of work So, what is the appeal of Portable Computer over just running Qwen 3.8 27B in Ollama? Everything around it. According to Perplexity, Portable Computer "packages step-level routing with the agent harness, local models, inference, tools, app connectors, and sandboxed execution in one maintained system." A good harness combined with the right tools and connectors is the difference between an AI model that is unwieldy and frustrating and one that is actually helpful. The stack around an LLM is a big part of what makes frontier models so compelling. That setup isn't exclusive to Portable Computer -- you could even build something similar yourself if you wanted to. The biggest advantage is convenience. A stack for a fully autonomous local AI, including skills and connectors, is a huge project. With Portable Computer, you can deploy the program on any Linux PC with the right hardware, though the current release targets the DGX Spark specifically. Windows support will arrive later this year. What do you need to run Portable Computer? Memory is the limit Portable Computer is designed to run using models with 27 billion parameters, and the official requirement is NVIDIA hardware with at least 32GB of VRAM. If you restrict the context or swap in a quantized model, you can get away with less. In the consumer card lineup, that means the RTX 3090 is basically the "weakest" GPU that you can use. Any RTX GPU with more VRAM will do fine; NVIDIA's DGX Spark, which features a whopping 128GB of RAM shared between the GPU and the CPU, is also up to the job. Whether you'll be able to use multi-GPU setups that add up to 24GB -- like two 16GB RTX 5060Tis -- is less clear. Perplexity says you can tweak the local stack to use your own inference server and models, which means it is probably possible, though it may require some tinkering to make it work. You can download and install Portable Computer today.

Perplexity's new local-first AI runs on your PC -- and asks before using the cloud

Portable Computer keeps your files and AI tasks on your own hardware, only turning to a cloud model when you give it permission Most AI assistants require you to send your questions, files and other data to somebody else's servers. Perplexity's new Portable Computer turns that arrangement around by running the entire AI system on hardware you own. Announced Tuesday, Portable Computer brings Perplexity's agentic Computer platform to Nvidia DGX Spark and other Linux machines with compatible Nvidia graphics cards. The company says every task begins on your machine. If the local model reaches a step that requires more advanced reasoning, Portable Computer asks for permission before sending that individual step to a frontier model in the cloud. The answer then returns to the local workflow, while files and other sensitive information can remain on the device. That could make it appealing to anyone who wants the convenience of an AI agent without automatically handing everything it touches to a cloud provider. There is, however, a fairly expensive catch. More than a locally installed chatbot You can already download open models and run them on your own computer. The difficult part is turning one of those models into an AI agent that can work with files, use tools and complete a task involving several steps. Doing that usually means finding and configuring an inference engine, agent harness, sandbox and any connectors needed to reach outside apps. You're also responsible for getting all those pieces to work together. Portable Computer packages that stack into one system. It includes the local model, inference engine, agent tools, app connectors and an isolated sandbox where it can execute code and use tools. The AI can work with files stored on your computer, read PDFs, run shell commands, take dictation and search the web. Users can also connect services including Google Drive, Gmail, Slack and GitHub. Code and tool use happen inside the local sandbox, according to Perplexity. That provides some separation between the AI agent and the rest of the machine, although I'd still want more detail about exactly what it can access and how those permissions are controlled. It knows when to ask for help The hybrid routing is probably the most interesting part of Portable Computer. A smaller local model may be enough to locate a file, extract information from a PDF or carry out a straightforward series of steps. It is less likely to match the reasoning capabilities of the largest models from OpenAI, Anthropic or Google. Perplexity's answer is to let the task remain local until the model encounters something it cannot reliably handle. Portable Computer can then ask whether it may send that step to a cloud model. Users don't have to select a privacy mode or create their own routing rules. The system makes the decision while it is working, but the user still has to approve the handoff. A supposedly local AI system becomes much less private if it quietly falls back to the cloud whenever a task gets difficult. Perplexity also says locally completed work won't consume billing credits. Only the portion sent to the cloud will count toward cloud usage. At launch, users can choose a 27-billion-parameter Qwen model or Perplexity's own post-trained version, called PPLX 27B. Nvidia's Nemotron 3.5 Lightning is expected to follow. Users will also be able to bring their own model and inference server. The hardware is the catch Portable Computer may run on your PC, but not just any PC. The software launches first on Linux and requires Nvidia hardware with at least 24GB of VRAM. It supports the desktop-sized Nvidia DGX Spark as well as compatible computers running DGX OS or Ubuntu on ARM or x64 hardware. Windows support is scheduled to arrive in September. Perplexity hasn't announced support for macOS. Portable Computer will be available to Perplexity Pro, Max, Enterprise Pro and Enterprise Max subscribers. Even with a compatible subscription, most people won't have the necessary hardware sitting under their desks. That makes this initial release more relevant to AI enthusiasts, developers and companies that already own high-end Nvidia systems than someone deciding between ChatGPT and Perplexity on an ordinary laptop. Still, the idea behind it could reach a much wider audience. AI assistants are gaining access to increasingly personal information, including email, work documents and files stored on our computers. Running more of that work locally gives users another way to decide what leaves their machine and when. Portable Computer isn't ready to replace a cloud chatbot for most people. It does, however, offer a glimpse of what a more private AI assistant could look like once the hardware needed to run one becomes easier to afford. Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom's Guide on YouTube and follow us on TikTok.

Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs

Perplexity is launching Portable Computer today, a version of its agentic "Computer" platform that runs entirely on hardware users already own -- starting with Nvidia's DGX Spark desktop supercomputer and Linux machines equipped with Nvidia RTX GPUs. The launch, developed in close partnership with Nvidia, is one of the most aggressive attempts yet to move serious AI agent workloads off the cloud and onto local devices. The model, the user's files, and the work itself can all stay on the machine. Work completed locally consumes no billing credits, and the company says every task starts on the device by default -- with the system asking permission before sending any individual step to a more powerful frontier model in the cloud. "We've basically brought the exact same UI to a fully local app," said Nate, Perplexity's vice president of engineering for infrastructure and enterprise, during a press briefing Monday. "This incorporates the entirety of the agent harness and inference and everything needed to do work locally." For Nvidia, which has spent the past two years selling the world on trillion-dollar AI data centers, the announcement signals something subtler but strategically important: the chipmaker believes local AI has crossed a threshold from hobbyist curiosity to practical tool -- and it wants to sell the hardware that runs it. "Local AI reached an inflection point," said Nader, Nvidia's director of developer technology, who focuses on developer tooling and open source. "For the longest time, it was hobbyists and enthusiasts, and they were running these quantized models that were quantized down to be super tiny... And while that's cool, it's not super practical. But all that changed with a lot of these new open source models that have come out that are super useful." How Portable Computer packages a full local AI stack into a single app Perplexity Computer, the company's agentic platform for knowledge work, orchestrates AI models, files, tools, and web access to complete multi-step tasks -- reviewing folders of documents, analyzing data, producing reports, and pushing results into business systems. Portable Computer replicates that experience locally: the local models, agent harness, inference engine, tools, app connectors, and a security sandbox come packaged together in a single system. That bundling is the point. With most local AI stacks today, users must assemble and operate those pieces separately -- downloading model weights, standing up an inference server, wiring together tools, and tuning performance. "Historically it's just been really painful to bring up the local AI stack," Nate said. "With Portable Computer, we really focused on just making this a really straightforward experience where you can get up and running very quickly." In one demo Monday, the system played the role of a retail investor reviewing a folder of 1099s and investment documents -- the kind of sensitive financial material many users would hesitate to upload to a cloud service. Running a 27-billion-parameter Qwen model at full GPU utilization on a DGX Spark, the agent reviewed each document and flagged cases where the hypothetical investor was paying unnecessary fees. The interface element that normally displays a running tally of cloud credits "is just parked at zero," Nate noted, "because all of this is happening on the device." A second demo showed the hybrid side of the product. Playing a startup founder, Nate asked the agent to analyze a CSV of user funnel data locally, then push the finished analysis to a Slack channel using Perplexity's connector ecosystem -- proof that local-first does not mean disconnected. The system also connects to Google Drive, Gmail, and GitHub, and can escalate to a frontier cloud model when the local model hits its limits. At launch, users can set up Qwen 3.8 27B or PPLX 27B, a version Perplexity has post-trained on its own harness, with Nvidia's Nemotron 3.5 Lightning coming soon. Portable Computer arrives today for Pro, Max, Enterprise Pro, and Enterprise Max subscribers on Linux, with Windows support following in September. Any RTX GPU with at least 24GB of VRAM -- roughly a GeForce RTX 3090 or newer -- clears the bar, a threshold Nate called "sort of the floor where we really want to make sure that we can deliver a great experience, but balance that with making it broadly available." Why co-designing the model and agent harness beats general-purpose frameworks Alongside the launch, Perplexity published a research paper arguing that effective local agents require the model and the agent harness -- the scaffolding of prompts, tools, and orchestration logic around the model -- to be designed together. The core insight: general-purpose harnesses assume a frontier model that can absorb enormous contexts, navigate sprawling tool surfaces, and plan over long horizons. Small local models buckle under those demands. Perplexity found empirically that although models like Qwen 3.8 27B advertise 260,000-token context windows, they begin to struggle beyond 100,000 tokens. So the company built a deliberately minimal harness: a succinct system prompt, a small set of core tools, and capabilities that load and unload as on-demand "skills" rather than sitting permanently in context. It converted popular connectors like Gmail and GitHub from token-hungry MCP servers into compact command-line tools, added self-verification hooks that monitor the health of a task, and enforced always-on OS-level sandboxing. If the sandbox is unavailable, the harness disables itself rather than running tools unprotected -- a contrast with open-source harnesses that run commands with the user's full permissions by default. The benchmark results Perplexity reports are striking, though they come from the company's own evaluations. On its internal Local Knowledge Work Bench -- 53 tasks spanning deep research, financial analysis, and document creation, which Perplexity says it plans to open-source -- Computer running Qwen 3.8 27B on a DGX Spark scored 82.6%, versus 77.6% for the open-source Pi harness and 74.0% for Hermes running the identical model. Perplexity's post-trained PPLX 27B pushed the score to 85.4%. The gaps widen dramatically on harder tasks: on BrowseComp, a web research benchmark, Computer hit 66.7% accuracy versus 50.2% for Pi and 43.9% for Hermes, while using 51% less wall time and 70% fewer tokens than Pi. On multimodal document understanding, Computer scored 65.1% against Hermes' 34.6% and Pi's 13.9%. The token economics driving AI agents from the cloud to local hardware The strategic logic behind the launch becomes clear when you consider how AI workloads have changed. Chat was bursty -- a question, an answer, done. Agents are different. "With agents, you want these agents always on if you can. You want the agents to really consume as many tokens as they can," Nader said. "What we're seeing is an insatiable demand for tokens, and that's something that makes local AI so great. As you saw through all these demos, you were not metered by the token. You were not paying for the token. So it's really killer for agents." This reframes the value proposition of local hardware. An agent that runs for hours reviewing documents, verifying its own work, and iterating on analyses would rack up substantial API bills in the cloud. On a device the user already owns, the marginal cost of those tokens approaches zero. Perplexity's paper makes the enterprise version of this argument explicitly: as agents scale across individual workflows and entire organizations, token expenditure and data movement "become increasingly difficult to govern." Local-first execution addresses both at once -- spend, because inference is free, and privacy, because sensitive tokens never leave the device boundary. Perhaps the most commercially interesting result concerns the hybrid middle ground. On Terminal Bench 2.1, a challenging coding benchmark, the fully local Qwen model scored 59.6% at essentially zero marginal cost. Letting it escalate to a Claude Opus 5 "advisor" in the cloud raised the score to 73.0% at an estimated $0.415 per task. Running the frontier model alone scored 82.4% at $0.65 per task. Escalation, in other words, recovered roughly three-fifths of the gap to frontier performance at about two-thirds of the cost -- and the user decides when that trade is worth making. Before any advisor call, the harness runs a PII classifier over the outgoing context and shows the user exactly what would leave the device. The remote model returns text guidance only; it never touches local files or tools. Where Portable Computer fits against Ollama and the DIY local AI stack Jason Hiner of The Deep View pressed the companies on how Portable Computer relates to existing local inference tools like Ollama. Nate's answer drew a clear line: the tools solve different layers of the problem. "The majority of the effort here has been at the agent harness level," he said, noting that the system uses vLLM to host model inference underneath, with an advanced mode for users who want to plug in their own inference endpoint. "We've heavily post-trained both the Qwen and Nemotron models that we're working with in order to really get the best possible results... Our focus has been on really honing the whole stack, top to bottom, of the model inference and the harness together." Nader put it more colorfully. "Just getting inference running really quickly on a Spark -- there's a smooth path. You can use Ollama. You can get that set up. But then, as you start to do more complicated, more agentic things, then suddenly you need more perf. You start looking at different models. You start looking at different harnesses, and it's kind of like the ocean. The deeper you go, the deeper it gets." The appliance-like pitch appeared to land with at least one attendee. Ben, who described struggling to set up his own DGX Spark despite being an engineer -- "this experience sucks, we have to fix it" -- said the product feels like the unlock "needed for people to really feel and understand what agentic means, and you need the right UX to make it happen." Nvidia also emphasized that the hardware scales: connecting two Sparks over shared memory runs frontier-class open models like DeepSeek's latest, and four can run GLM 5.2 or Nemotron Ultra. "I've even seen eight Sparks get connected," Nader said. What the deepening Nvidia-Perplexity alliance means for both companies The launch extends a partnership that has been building for more than a year. In June 2025, Nvidia and Perplexity announced a collaboration to bring sovereign AI models to European publishers and telecoms, part of CEO Jensen Huang's continent-hopping campaign to convince governments that, as the Associated Press reported from VivaTech in Paris, "every country needs a national intelligence infrastructure." The sovereign AI pitch -- that data "belongs to your people, your country, your culture," in Huang's words -- is philosophically the same argument Portable Computer makes at the scale of a single desk: intelligence you control, running on hardware you own. There is a self-interested logic for both companies. Perplexity, which has raised capital at steadily escalating valuations while facing legal pressure from publishers over its content practices -- including a lawsuit filed by The New York Times in December 2025 and an earlier public dispute with Forbes -- gets a product whose economics don't depend on metering every token, and a differentiated wedge into privacy-sensitive enterprises in law, healthcare, and finance. Nvidia gets a killer app for DGX Spark, a device that, by the admission of attendees at Monday's briefing, has been easier to buy than to use. When one reporter asked whether a Spark might ship with Portable Computer and a Nemotron model preinstalled, Nader demurred without ruling it out: "That would be cool... the goal is just making sure that it's a super smooth experience for every user." Questions remain. Perplexity's most impressive numbers come from its own internal benchmark, and the company acknowledges that compact models still trail the frontier meaningfully on hard reasoning tasks -- advisor escalation "narrows but does not fully close the gap." The launch is Linux-only for now, the 24GB VRAM floor excludes the vast majority of consumer PCs, and Apple silicon -- home to some of the most enthusiastic local AI tinkerers -- is conspicuously absent from the roadmap. "We're very focused right now on Nvidia hardware," Nate said when asked. But the direction of travel is unmistakable. Perplexity's researchers describe the launch as part of "a broader shift in which increasingly capable agents move from remote infrastructure to individual and local devices," and both companies are betting that advances in chips and open models will keep expanding what a box on a desk can do. During Monday's demos, the most telling detail wasn't a benchmark score -- it was that credit counter in the corner of the screen, sitting motionless at zero while the agent churned through a folder of tax documents. For two years, the AI industry has measured its ambitions in gigawatts and tokens per dollar. Portable Computer proposes a different meter, one that never runs.

Perplexity AI launches Portable Computer on-device AI agent

Perplexity AI launches Portable Computer on-device AI agent Perplexity AI Inc. today introduced Portable Computer, an artificial intelligence agent designed to run on desktops equipped with Nvidia Corp. silicon. The launch follows a report that Nvidia is weighing an investment in the startup that could value it at over $30 billion. Furthermore, Nvidia has reportedly floated the idea of licensing Perplexity's technology and hiring key employees. Portable Computer is an on-device version of a cloud-based AI agent called Perplexity Computer that debuted in February. Both tools are designed to perform multi-step tasks such as developing web applications. When Perplexity Computer encounters a particularly time-consuming task, it split the work into smaller chunks and assigns each one to a different subagent. Portable Computer brings similar capabilities to Nvidia's DGX Spark device (pictured.) It's a desktop computer that features a graphics card based on the company's Blackwell architecture. Additionally, the system ships with a 20-core central processing unit and 128 gigabytes of memory. On launch, Portable Memory uses the open-source Qwen 3.8 27B language model to automate tasks. It also supports a version of the algorithm called PPLX 27B that Perplexity optimized for the DGX Spark. The company added in a so-called multi-token prediction mechanism designed to speed up prompt processing. Qwen 3.8 27B has a context window of 256,000 tokens. In practice, however, Perplexity determined that the model struggles to process prompts with more than 100,000 tokens. The company addressed that limitation by equipping Portable Computer with a so-called context compaction feature. It summarizes lengthy prompts to bring them below the 100,00 token mark. Personal Computer includes several pre-packaged skills, file collections that include instructions and other assets designed to boost the quality of prompt responses. Perplexity says that the initial skill lineup focuses on use cases such as online research, data science and coding. Additionally, the company has added in a mechanism that verifies the accuracy of prompt responses. If Portable Computer's on-device model struggles to complete a task, it can request help from a cloud-based neural network. It can also connect to several popular software-as-a-service applications including GitHub. Portable Computer requests user permission before sending data to external tools and doesn't provide them with access to on-device files. According to Perplexity, the software also includes other guardrails. It features a sandbox that prevents language models from accessing parts of the user's operating system that aren't relevant to their work. Additionally, Portable Computer blocks unauthorized network connections. An upcoming update will expand the tool's availability beyond the DGX Spark to Windows machines that feature Nvidia RTX graphics cards. Additionally, Perplexity plans to add support for the chipmaker's recently released Nemotron 3.5 Lightning model. It's a mixture of experts algorithm with 30 billion parameters that is optimized to run on devices with limited processing capacity.