PPLX 27B Launches in Perplexity's Portable Computer: A Local Agent That Beats the Open Harnesses

Perplexity's PPLX 27B is not a cloud model, and that is the entire story. Announced August 25, 2026, alongside Portable Computer — the company's local-first agent, built with NVIDIA — PPLX 27B is a post-trained version of Alibaba's open-weight Qwen 3.8 27B, tuned by Perplexity for its own agent harness and shipped to run entirely on hardware you own. On Perplexity's 53-task Local Knowledge Work Bench, the harness running PPLX 27B scored 85.4%, ahead of the same harness on Qwen 3.8 27B (82.6%) and of two open-source agent harnesses, Pi (77.6%) and Hermes (74.0%). Those are the vendor's own numbers from the launch post and the companion research paper, "A Local-First Agent for Private and Cost-Effective Knowledge Work," not independently reproduced — but they are the first public evidence that a 27B model on a desktop can carry real knowledge work at a fraction of the cloud's price.

What actually shipped

Portable Computer is the on-device build of Perplexity's "Computer" agent platform. Where the earlier Computer product ran its orchestrator in the cloud, Portable Computer packages the whole stack into one system that runs on your machine: the agent harness, orchestrator, planner, tool router, scheduler, a durable task queue, a local search index, the inference engine, and the app connectors for Google Drive, Gmail, Slack, GitHub, and Outlook.

Two properties separate it from a DIY local setup. First, code and tool execution happen inside an OS-enforced sandbox, and if the sandbox is unavailable, tool execution is disabled rather than run unprotected. Second, it is local-first by design: every task starts on-device, and if a step needs live web data or frontier reasoning, the orchestrator stops and asks for permission before escalating that single step to one of 15+ cloud models. A PII classifier runs over the outgoing context first and shows you exactly what would leave the machine.

The two 27B models

At launch you choose between two 27-billion-parameter models. Qwen 3.8 27B is Alibaba's Apache-2.0 open-weight model — a dense, multimodal 27B with a 262K-token native window, released two weeks earlier and already a strong self-host option. PPLX 27B is that same base post-trained by Perplexity: not a new architecture, but additional training tuned specifically for this harness, with the company claiming more precise and more efficient behavior on its own agent workloads. NVIDIA's Nemotron 3.5 Lightning (30B) is listed as coming soon, and bring-your-own model and inference server are both supported, so the harness is not married to Perplexity's weights.

The benchmark, and what it is not

The headline figure comes from the 53-task Local Knowledge Work Bench, a suite spanning the sort of work a knowledge worker would actually delegate — deep research, financial analysis, document creation. Perplexity says it plans to open-source the benchmark, which will eventually make the comparison repeatable rather than trusted on faith. Until then, these results are vendor-run. On that suite, per the vendor:

• Portable Computer with PPLX 27B — 85.4%

• Portable Computer with Qwen 3.8 27B — 82.6%

• Pi (open-source harness) — 77.6%

• Hermes (open-source harness) — 74.0%

Two more vendor-reported results round out the picture. On BrowseComp, the web-research suite, Computer scored 66.7% against 50.2% for Pi and 43.9% for Hermes, while using 51% less wall time and 70% fewer tokens than Pi. On ParseBench-100, a document-extraction test, it scored 65.1% against 34.6% for Hermes and 13.9% for Pi. The widest gaps land on the tasks that reward a harness's own infrastructure — search, routing, tool use — which is where a post-trained model plus a purpose-built plumbing stack earns its keep.

The economics flip

The more interesting number is the price. Work completed locally carries zero per-token cost and consumes no Perplexity credits — the counter sits at zero for a fully local task. Escalation is the only thing that costs money, and it is opt-in per step. Perplexity published the hybrid math on Terminal Bench 2.1: fully local scored 59.6% at roughly zero cost; adding a frontier model as an advisor lifted it to 73.0% at about $0.415 per rollout; the frontier model alone reached 82.4% at about $0.65 per rollout. That last contrast is the thesis — local inference flattens the cost curve for always-on agents, and cloud escalation becomes a surgical expense where it pays for itself.

None of which is free to start. The box matters. Portable Computer launches Linux-first on the NVIDIA DGX Spark, the desktop supercomputer with 128GB of unified memory (list price around $4,500), or on Linux PCs with RTX GPUs of at least 24GB VRAM — roughly an RTX 3090 or newer — with 32GB recommended. Windows support is scheduled for September; macOS is not announced. And the software itself requires a paid Perplexity plan: Pro, Max, Enterprise Pro, or Enterprise Max. This is a subscription-plus-hardware product aimed at people who already pay Perplexity, not a general API.

Where the router fits

PPLX 27B is not something OrcaRouter routes — it runs on hardware you own, behind a Perplexity subscription, and we won't pretend otherwise. What we do route is the comparison set this launch invites. Qwen 3.8 27B, the exact base weights PPLX 27B post-trains from, is live on OrcaRouter self-hosted; so are DeepSeek V4 Flash, Gemini 3.5 Flash Lite, and GPT-5.6 Luna — all behind one API key at provider list price, passed through with zero markup, so a vendor price cut is live on our side the same day it is announced.

That matters here for one specific reason. Portable Computer's pitch is the local-versus-cloud cost contrast, and the cloud half of that contrast is only as honest as the prices you compare against. On OrcaRouter those are the provider's actual list prices, which turns the local-or-cloud decision into a calculation you can trust rather than a marketing comparison. And if you want to prototype the same agent both ways — local harness on one box, cloud models behind one endpoint — automatic failover lets the cloud side stand in when the local model is out of depth, without a second contract or a second code path.

What to watch

Three things over the next month decide whether this is a niche product or the start of a category. Whether Perplexity actually open-sources the Local Knowledge Work Bench, which would turn the headline numbers from claims into something the community can reproduce. Whether Nemotron 3.5 Lightning lands and beats PPLX 27B on the same harness — the "post-train for the harness" thesis is only as good as the base model beneath it. And whether Windows support in September broadens the reach beyond DGX Spark owners. If the benchmark is real and the harness transfers, "run the agent on your own GPU" stops being a hobbyist pattern and becomes a deployment option with real teeth — and the cloud models it gets compared against will have to earn their per-token prices.