Perplexity has introduced Portable Computer, an AI agent framework designed for on-device models like Qwen 3.8 27B, enabling zero-cost inference. The framework runs locally, calling external services only when necessary, thus preserving privacy and reducing costs. In benchmarks, it outperformed general-purpose frameworks, showcasing the potential of local AI.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Perplexity introduces Portable Computer, a local-first agent framework built around the on-device model Qwen 3.8 27B, claiming near-zero inference costs for knowledge work while protecting privacy. This article, based on the official blog and a report by Machine Intelligence, breaks down its design principles, benchmark data, and limitations.
Perplexity introduces Portable Computer, aiming to leverage locally deployable models (such as Qwen 3.8 27B) to achieve results close to cloud-based large models. The core motivation is to address token consumption and data privacy issues: closed-source model APIs take private information off-device with every request, whereas local inference incurs no fees and data stays within the device boundary.
The framework design follows several key principles: context efficiency (keeping the core framework concise, loading skills on demand, supporting context compression), connectors as command-line tools (avoiding the large tool definitions of MCP servers consuming context), self-verification (model or hooks trigger verification to improve performance), and sandbox execution (OS-level sandbox restricting processes, file paths, and network access, enabled by default and not degradable).
The coordinator is deterministic framework code, not a large model, responsible for maintaining the loop, building context, and executing policies; the local model proposes next actions, and the coordinator executes them in the sandbox. External functions (such as web search, connectors, advisors) only cross the device boundary when enabled and approved by the user.
To evaluate the framework's performance in isolation, Perplexity used the same Qwen 3.8 27B model (not post-trained) on DGX Spark to compare Portable Computer with open-source frameworks Hermes and Pi. In web search tests based on 1266 BrowseComp tasks: Computer achieved 66.7% accuracy, Pi 50.2%, Hermes 43.9%; average runtime 402.1 seconds, 852k tokens, 61% less time and 16% fewer tokens than Hermes, and 51% less time and 70% fewer tokens than Pi.
In multimodal document understanding on ParseBench-100 (100 tasks, covering charts, layouts, tables, text, and formatting, 20 each): Computer averaged 65.1%, Hermes 34.6%, Pi 13.9%; Computer took 60.6 seconds per task and 20.1k tokens, far lower than Hermes (108.3 seconds, 32.1k) and Pi (410.5 seconds, 829.1k). Computer led in all five categories, but layout remains a common challenge for all three.
For complex tasks beyond the capabilities of on-device models, the framework provides an advisor tool: the local model can request a more powerful frontier model (such as Claude Opus 5) for planning, ambiguity resolution, failure recovery, or result verification. Escalation is optional, and users control how it is enabled and approved; the advisor only receives approved context, cannot directly access device files or tools, and a PII classifier flags sensitive information before sending.
On Terminal Bench 2.1 (89 coding tasks), the fully local Qwen 3.8 27B scored 59.6%, improving to 73.0% with the Claude Opus 5 advisor, with an API cost of about $0.415 per deployment; running Claude Opus 5 alone scored 82.4% at a cost of $0.65. Escalation closes three-fifths of the gap at about two-thirds of the frontier cost, but does not fully eliminate it.
Perplexity also post-trained Qwen 3.8 27B using synthetic data (rejection sampling fine-tuning + reinforcement learning), producing PPLX 27B. On the internal Local Knowledge Work Bench (53 tasks covering seven knowledge work categories), Computer with the base model scored 82.6% (Pi 77.6%, Hermes 74.0%), and PPLX 27B improved the score to 85.4%, but with increased token consumption and an estimated runtime of 250 seconds.
Despite the clear advantages of the local-first approach in cost and privacy, on-device models still fall short of frontier models in absolute performance. On the high-difficulty coding tasks of Terminal Bench 2.1, local models lagged behind frontier models on all test platforms, and advisor escalation could only narrow the gap, not eliminate it.
Perplexity believes that as chips, models, and devices advance, the scope and quality of knowledge work handled locally will continue to expand. They plan to release a technical report detailing the training process and open-source the evaluation benchmarks.
This article is primarily based on Perplexity's official blog (reproduced by Machine Intelligence), which is self-reported. All benchmarks were designed or selected by Perplexity, so there may be a conflict of interest. Some data (such as BrowseComp, ParseBench-100, Terminal Bench 2.1) come from third-party benchmarks, but the test conditions were controlled by Perplexity. The PPLX 27B scores come from an internal benchmark and have not been independently verified.
Perplexity Portable Computer demonstrates the feasibility of local-first agents: through a carefully designed framework and post-training, on-device models can approach frontier model performance in knowledge work scenarios while significantly reducing costs and protecting privacy. However, high-difficulty tasks still require cloud-based advisors, and all performance data comes from Perplexity itself, so it should be interpreted with caution.
Primary report
Primary source