# Perplexity Portable Computer

> Source: https://aiwiki.ai/wiki/perplexity_portable_computer
> Updated: 2026-08-25
> Fact-checked: 2026-08-25
> Categories: AI Agents, AI Tools & Products, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Perplexity Portable Computer." aiwiki.ai, 25 Aug 2026. https://aiwiki.ai/wiki/perplexity_portable_computer
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Perplexity Portable Computer** is a local-first [AI agent](https://aiwiki.ai/wiki/ai_agents) application released by Perplexity AI on August 25, 2026, in collaboration with [NVIDIA](https://aiwiki.ai/wiki/nvidia). It is a software version of the company's Perplexity Computer agent that runs its model, agent harness, task state, and tool-execution loop on hardware controlled by the user. At launch, Perplexity made it available to Pro and Max subscribers using Linux on an [NVIDIA DGX Spark](https://aiwiki.ai/wiki/nvidia_dgx_spark).[1][4][7]

The name does not refer to a laptop or to the general category of portable computers. The launch product was an application and model stack for an existing desktop AI system. Perplexity described it as local-first rather than local-only: a task starts on the device, but web search, connected services, and stronger cloud models can be invoked with user authorization. Work completed by a local model does not consume Perplexity credits, while a cloud handoff can use credits or incur an API cost. Hardware, electricity, and subscriptions remain outside that narrow billing claim.[1][2][8]

## Key facts

| Field | Detail |
| --- | --- |
| Developer | Perplexity AI, working with NVIDIA on the launch configuration[1][4] |
| Released | August 25, 2026[1][7] |
| Product type | Local-first agent application and execution stack[1][2] |
| Launch platform | NVIDIA DGX Spark running Linux[1][3] |
| Launch access | Perplexity Pro and Max subscribers[1][3] |
| Local components | Model inference, harness, deterministic orchestrator, planner, tool router, scheduler, durable task queue, local search index, conversation state, and task trajectory[1][2] |
| Launch models | Perplexity's announcement named Qwen 3.8 27B and PPLX 27B; its setup page also listed Qwen 3.6 35B[1][3] |
| Optional off-device services | Perplexity Search, connected applications, web access, and cloud advisor models, subject to user-controlled approval settings[1][2] |
| Initial connectors | Google Drive, Gmail, Slack, and GitHub; the setup page also named Outlook[1][3] |
| Announced future support | Windows, GeForce RTX and RTX PRO PCs, DGX Station, and a Nemotron 3.5 Lightning option[1][4] |
| Public implementation status | No public source repository or software license for the application or harness was located at launch; a related Qwen serving artifact used different access and license terms[2][6] |
| Evaluation status | Perplexity published internal and public-benchmark results; no independent end-to-end replication was located by August 25, 2026[2][7] |

## Product design

Portable Computer packages more than [large language model](https://aiwiki.ai/wiki/large_language_model) inference. Its local execution loop includes an agent harness, a planner, tool selection, sandboxed execution, task scheduling, local indexing, and persistent state. Perplexity's technical description says a deterministic orchestrator, rather than another language model, maintains that loop. The local model proposes the next action, the orchestrator checks policy and executes approved tool calls, and the result returns to the model as new context.[2][3]

This division matters because a model and an agent system are not interchangeable. A model produces candidate text or actions, while the surrounding harness determines which tools are visible, how files are presented, when context is shortened, what commands can run, and when a remote service may be consulted. Consequently, the product's evaluation results describe model, harness, search provider, and execution-policy combinations. They are not scores for Qwen or PPLX 27B in isolation.[2]

Perplexity designed the harness around the limits it observed in compact local models. Although Qwen 3.8 27B has a nominal 260,000-token [context window](https://aiwiki.ai/wiki/context_window), the company said its own testing found reliability began to decline beyond about 100,000 tokens. Portable Computer therefore uses a short core system prompt and a limited set of core tools. Other capabilities load as skills only when needed, and stale trajectory material can be compacted when a task becomes long.[2]

The company also said it converted frequently used [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol) servers into command-line tools with smaller definitions. A conventional MCP server may expose enough tool schema to occupy a meaningful part of a local model's context before any user data is considered. Perplexity's approach reduces that fixed tool-description cost, but the company did not publish a controlled ablation showing how much each conversion contributed to final task quality.[2]

Self-verification is another part of the harness. The model may check its own work, while trajectory-monitoring hooks can request a verification step after a failure or unhealthy run. These checks consume additional local tokens and time. Perplexity reported higher final-task scores when verification was used, but did not release a component-by-component study that separates verification, context compaction, tool packaging, search infrastructure, and prompting.[2]

## Local execution and cloud escalation

Every task begins locally according to Perplexity's launch design. The model, harness, conversation, and trajectory remain on the device by default. Local file reading and document processing can therefore occur without sending those inputs to a remote model. The product page also describes a local search index, a durable task queue, and the ability to retain state across sessions.[1][2][3]

Tasks can still cross the device boundary. Web search, Gmail or GitHub actions, other connectors, and a cloud advisor necessarily involve external services. For an advisor call, Perplexity says the orchestrator selects the relevant context, applies a personally identifiable information classifier, and shows the material proposed for transmission. The remote model returns text advice and receives no direct authority over device files or tools. The local model and orchestrator continue the task after the advice returns.[2]

Approval can be configured in more than one way. The technical article says a user may approve every advisor call manually or enable automatic approval. The latter is still a user setting, but it is not a per-call confirmation. This makes "asks before using the cloud" an accurate description of the default interaction presented at launch only when approval has not already been granted more broadly. A classifier or approval dialog can reduce unintended disclosure, but neither independently proves that a selected context is nonsensitive.[1][2][9]

A user can avoid cloud-model escalation and disable web search for a fully local workflow, according to Perplexity. That mode also gives up current web information and connected-service actions. The product is therefore a hybrid between [offline inference](https://aiwiki.ai/wiki/offline_inference) and [cloud computing](https://aiwiki.ai/wiki/cloud_computing), with the boundary controlled through enabled capabilities and approval policy rather than by a claim that the machine never communicates externally.[2][8]

## Models and post-training

Perplexity's launch article named two local choices: the base [Qwen3.8](https://aiwiki.ai/wiki/qwen3_8) 27B model and PPLX 27B, which it described as a post-trained version of Qwen 3.8 27B. The separate setup page listed Qwen 3.8 27B and Qwen 3.6 35B, while labeling NVIDIA Nemotron 3.5 Lightning as forthcoming. Those launch surfaces did not present an identical model menu, so Qwen 3.6 35B and PPLX 27B should not be treated as alternate names for the same model.[1][3]

For PPLX 27B, Perplexity said it first identified representative categories from Computer usage, then generated synthetic tasks with containerized environments and automatic verifiers. It described a two-stage process: rejection fine-tuning on high-scoring trajectories followed by [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning). The company said the synthetic environments contained no real documents or user information. A held-out set of 53 tasks became its Local Knowledge Work Bench.[2]

Important training details were still missing at launch. Perplexity promised a technical report and said it planned to open the internal benchmark, but neither was available by the cutoff. The publication did not disclose a complete training corpus, training compute, hyperparameters, contamination analysis, model weights, or a release-specific license for PPLX 27B. The internal benchmark's environments and verifiers also could not be independently inspected.[2]

A Perplexity Hugging Face page did document a quantized Qwen 3.8 27B checkpoint used in a Computer deployment on DGX Spark. Its card described NVFP4 weights, FP8 key-value cache, speculative decoding, patched vLLM serving files, and a configured 262,144-token maximum. The page labeled the license as `other`, and its instructions required a Perplexity read token. It did not identify that artifact as PPLX 27B, so it is evidence about a serving configuration, not a public release of the post-trained model or the application.[6]

## Hardware and installation

The first supported system was NVIDIA DGX Spark, a compact [NVIDIA Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) computer built around the GB10 Grace Blackwell platform. NVIDIA lists a 20-core Arm CPU, a Blackwell GPU, and 128 GB of coherent unified memory. Portable Computer's setup page specified a DGX Spark with 128 GB of memory and at least 1 TB of storage. These specifications describe the launch platform's capacity; NVIDIA's precision-dependent peak compute figure is not an end-to-end Portable Computer performance measurement.[3][5]

Perplexity described installation as one-click through its app, while its public setup page also provided commands that add a Perplexity APT repository and install a Linux package. The page says setup installs the local harness, orchestrator, models, and sandboxed execution. Model downloads and first startup can still take time, and the product requires enough memory and storage for the selected local model.[1][3]

NVIDIA and Perplexity announced broader hardware support rather than shipping it on launch day. NVIDIA said GeForce RTX and RTX PRO support was coming, while Perplexity named Windows and DGX Station as future targets. The launch article's phrase "available on Linux" should not be generalized to every Linux PC without a documented compatibility list. Independent reports sometimes described compatible NVIDIA Linux machines more broadly, but the first-party setup path and availability statement centered on DGX Spark.[1][4][9]

The hardware changes the product's economics but does not eliminate cost. Local inference avoids a metered model API for work kept on the machine, yet it requires the device, power, storage, maintenance, and a qualifying Perplexity plan. Computerworld quoted analysts who viewed the specialized hardware investment as a meaningful barrier. An independent DGX Spark review published before Portable Computer likewise found that the system could run substantial local workloads but retained Linux, model-download, and setup friction.[8][17]

## Provider-run evaluations

Perplexity evaluated the local harness on one internal task set and three public benchmarks. Unless otherwise stated, its comparisons used the same Qwen 3.8 27B model and a DGX Spark, then changed the harness. The reported figures came from Perplexity and had not been independently reproduced by the cutoff.[2][7]

| Evaluation | Scope | Perplexity-reported result | Main interpretation limit |
| --- | --- | --- | --- |
| Local Knowledge Work Bench | 53 held-out tasks in seven knowledge-work groups, three trials per task | Computer with Qwen 3.8 27B, 82.6%; Pi, 77.6%; Hermes, 74.0%; Computer with PPLX 27B, 85.4% | Internal tasks, environments, and verifiers were not public[2] |
| BrowseComp | 1,266 web-research questions | Computer, 66.7%; Pi, 50.2%; Hermes, 43.9% | Computer used Perplexity search, while the other harnesses used Brave[2][10] |
| ParseBench-100 | 100 document tasks, 20 in each of five categories | Computer, 65.1%; Hermes, 34.6%; Pi, 13.9% | Perplexity evaluated a subset rather than the roughly 2,000-page full benchmark[2][11] |
| Terminal-Bench 2.1 | 89 terminal tasks | Local Qwen, 59.6%; Qwen with Claude Opus 5 advisor, 73.0%; Claude Opus 5 alone, 82.4% | All three used the Computer harness; no equivalent Pi or Hermes advisor condition was tested[2][12][13] |

The Local Knowledge Work Bench includes research, finance and data work, document production, engineering, compliance, dashboards, and project tasks. Perplexity reported that the base Qwen configuration used an average of 520,000 tokens and 218 seconds per task, compared with 681,000 tokens and 176 seconds for Pi and 634,000 tokens and 292 seconds for Hermes. PPLX 27B raised the reported score while increasing the averages to 678,000 tokens and 250 seconds. This means the best reported score was not also the fastest or least-token configuration.[2]

On BrowseComp, Perplexity reported 402.1 seconds and 852,000 tokens per task for Computer, compared with 1,020.9 seconds and 1.01 million tokens for Hermes and 826.0 seconds and 2.82 million tokens for Pi. The benchmark's original paper defines it as 1,266 short-answer questions that require persistent browsing, while cautioning that it does not reproduce the ambiguity and long-form outputs of ordinary user work. Perplexity's differing search backends add another confound: the result tests both harness and retrieval system, not only agent orchestration.[2][10]

For ParseBench-100, the company reported averages of 60.6 seconds and 20,100 tokens for Computer, 108.3 seconds and 32,100 tokens for Hermes, and 410.5 seconds and 829,100 tokens for Pi. The full ParseBench paper covers roughly 2,000 human-verified enterprise-document pages and more than 169,000 rules across tables, charts, content faithfulness, formatting, and visual grounding. Perplexity's 100-task evaluation was balanced across five categories, but its article did not provide a complete subset-construction and selection audit.[2][11]

The Terminal-Bench experiment measured the product's optional advisor design. Perplexity estimated $0.415 in advisor API cost per rollout for the hybrid condition and $0.65 for Claude Opus 5 alone. These are provider estimates for model calls, not total cost of ownership. The comparison does not include the DGX Spark, electricity, subscription, or local-model opportunity cost. It also does not isolate the advisor from the Computer harness because all conditions used that harness and neither Pi nor [Hermes Agent](https://aiwiki.ai/wiki/hermes_agent) was modified to expose an equivalent advisor tool.[2]

The benchmark publications help define the tasks but do not validate Portable Computer. BrowseComp is an OpenAI-authored benchmark for difficult web lookup. ParseBench measures document-parsing dimensions. Terminal-Bench uses containerized command-line tasks with human-written solutions and verification tests, while version 2.1 modified numerous tasks for bugs, resources, timeouts, or reward-hacking robustness. Confidence intervals reported by Perplexity quantify sampling uncertainty under its runs; they do not remove systematic differences in tools, search providers, prompts, subset choice, or unpublished evaluation assets.[2][10][11][12][13]

## Security and privacy boundary

Perplexity says Portable Computer executes code and tools in an operating-system-level sandbox that restricts processes, filesystem paths, and network access according to policy. If that sandbox is unavailable, the harness disables tool use rather than running commands with ordinary user permissions. These are meaningful design claims, but the launch materials do not publish the exact local sandbox engine, policy defaults, penetration-test results, patch process, or a product-specific threat model.[1][2]

Perplexity separately documented SPACE, the virtual-machine sandbox platform used for its broader Computer service. SPACE includes guest-kernel isolation, network egress control, credential management, snapshots, and a host-to-guest daemon. The Portable Computer documents do not state that the local product uses the same VM backend or credential path. SPACE therefore provides company architecture background, not proof of Portable Computer's exact implementation.[16]

The distinction between local-first and local-only is also security-relevant. An approved advisor request transmits selected context to a cloud model. Connectors can send or modify information in third-party services, and web search discloses a query to remote infrastructure. Perplexity's PII classifier and preview step are controls around those transfers, not guarantees that every trade secret, contractual restriction, or context-dependent sensitive fact will be recognized.[1][2][8]

Computerworld reported that enterprise security specialists wanted controls stronger than a user prompt, including a policy that could forbid cloud requests instead of asking users to decide repeatedly. The article raised prompt-manipulation and reflexive-approval scenarios as concerns, not as documented exploits. Tom's Guide similarly noted that more detail was needed about the sandbox and permission surface. No independent security audit of Portable Computer was located by August 25.[8][9]

Academic work supports a cautious interpretation without evaluating this product. Privacy-R1 treats local-versus-cloud delegation as a sequential tradeoff between task utility and information disclosure. A separate 2026 comparative study of AI code sandboxes emphasizes that the deployable product, underlying isolation engine, and patch policy are distinct evidence layers, and that no single architectural property establishes overall security. Neither study tested Portable Computer, its PII classifier, its approval interface, or its sandbox.[14][15]

## Availability, licensing, and reproducibility

At launch, Portable Computer was distributed through Perplexity to Pro and Max subscribers rather than as a standalone open-source project. The setup page made the Linux package path and model-download process visible, but no public source repository or software license was located for the agent harness, orchestrator, sandbox policy, connectors, or application interface. The absence of a located repository is not evidence about whether Perplexity may release components later.[2][3]

Reproducing the published results would require more than the application package. The Local Knowledge Work Bench was not available, PPLX 27B lacked its promised technical report, and the exact prompts, task images, evaluation containers, verifier code, and all model artifacts were not published together. The public Qwen serving card documents one inference configuration, but it is access-controlled, carries an `other` license label, and is not the post-trained PPLX 27B release.[2][6]

Independent coverage confirmed that the product existed and accurately identified its local-first goal, but the detailed architecture, security behavior, and performance figures remained primarily vendor-supplied. Later publication of the benchmark, model report, harness source, local sandbox design, or an independent end-to-end audit would materially change the product's reproducibility and evidence status.[7][8][9]

## References

1. Perplexity AI. "Introducing Portable Computer for local-first AI." Aug. 25, 2026. https://www.perplexity.ai/hub/blog/introducing-portable-computer-for-local-first-ai
2. Perplexity AI. "A Local-First Agent for Private and Cost-Effective Knowledge Work." Aug. 25, 2026. https://www.perplexity.ai/hub/blog/a-local-first-agent-for-private-and-cost-effective-knowledge-work
3. Perplexity AI. "Portable Computer on NVIDIA DGX Spark." Accessed Aug. 25, 2026. https://pplx-dgx-spark-lp-demo.pplx.app/
4. NVIDIA. "NVIDIA and Local AI Community Fuel Open Source Models and Intelligent Agents." Aug. 25, 2026 update. https://blogs.nvidia.com/blog/local-ai-open-source-models-agents-nemotron/
5. NVIDIA. "NVIDIA DGX Spark Specifications." Accessed Aug. 25, 2026. https://www.nvidia.com/en-us/products/workstations/dgx-spark/
6. Perplexity AI. "pplx-qwen-3-8-27b-dflash2-20260819." Hugging Face model repository, accessed Aug. 25, 2026. https://huggingface.co/perplexity-ai/pplx-qwen-3-8-27b-dflash2-20260819
7. Michael Nunez. "Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs." VentureBeat, Aug. 25, 2026. https://venturebeat.com/infrastructure/perplexity-partners-with-nvidia-to-launch-portable-computer-a-fully-local-ai-agent-with-zero-token-costs
8. Evan Schuman. "Perplexity's on-device AI offering promises data control and lower token costs." Computerworld, Aug. 25, 2026. https://www.computerworld.com/article/4213821/perplexitys-on-device-ai-offering-promises-data-control-and-lower-token-costs.html
9. Amanda Caswell. "Perplexity's new local-first AI runs on your PC and asks before using the cloud." Tom's Guide, Aug. 25, 2026. https://www.tomsguide.com/ai/perplexitys-new-local-first-ai-runs-on-your-pc-and-asks-before-using-the-cloud
10. Jason Wei et al. "BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents." arXiv:2504.12516, 2025. https://arxiv.org/abs/2504.12516
11. Boyang Zhang et al. "ParseBench: A Document Parsing Benchmark for AI Agents." arXiv:2604.08538, 2026. https://arxiv.org/abs/2604.08538
12. Mike A. Merrill et al. "Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces." arXiv:2601.11868, 2026. https://arxiv.org/abs/2601.11868
13. Harbor Framework. "Terminal-Bench 2.1." GitHub repository, accessed Aug. 25, 2026. https://github.com/harbor-framework/terminal-bench-2-1
14. Zheng Hui et al. "Privacy-R1: Privacy-Aware Multi-LLM Agent Collaboration via Reinforcement Learning." ACL 2026. https://aclanthology.org/2026.acl-long.2130/
15. George Andronchik and Pavel Lokhmakov. "AI Code Sandboxes: A Comparative Security Study. Part 1 of 2." arXiv:2606.08433, 2026. https://arxiv.org/abs/2606.08433
16. Perplexity AI. "Making SPACE: Secure and efficient runtimes for long-running agents." July 15, 2026. https://www.perplexity.ai/hub/blog/making-space-secure-and-efficient-runtimes-for-long-running-agents
17. Mithun Mohandas. "NVIDIA DGX Spark review: This is what a real AI PC feels like." Digit, June 24, 2026. https://www.digit.in/features/general/nvidia-dgx-spark-review-this-is-what-a-real-ai-pc-feels-like.html