Perplexity & Nvidia Launch "Portable Computer"
Local-First AI Agents — Zero Token Costs
The Big News
Perplexity today launched **Portable Computer**, a fully local version of its agentic "Computer" platform, built in partnership with **Nvidia**. It runs on hardware users already own — starting with the **DGX Spark** and Linux machines with RTX GPUs (24GB+ VRAM, e.g., RTX 3090). Windows support arrives in September.
Everything — model, files, agent harness, tools, security sandbox — runs on-device. Work completed locally costs **zero credits**, and the system asks permission before escalating any step to a cloud frontier model.
> *"We've basically brought the exact same UI to a fully local app."* — Nate, VP of Engineering, Perplexity
>
> *"Local AI reached an inflection point."* — Nader, Director of Developer Technology, Nvidia
What It Is
A single packaged app that replaces the painful DIY local stack (model weights + inference server + tools + tuning):
- Bundles models, inference engine, agent harness, connectors (Google Drive, Gmail, GitHub, Slack), and OS-level sandboxing
- Models at launch: **Qwen 3.8 27B** or **PPLX 27B** (post-trained by Perplexity); Nemotron 3.5 Lightning coming soon
- Hybrid mode: escalate to frontier cloud models when local limits are hit — user approves every step; a PII classifier screens outgoing context; remote models return text guidance only
Key Insight: Co-Design Model + Harness
Perplexity's research paper argues small local models fail under general-purpose agent scaffolds. Their minimal harness uses short prompts, few core tools, on-demand "skills," CLI-based connectors instead of token-heavy MCP servers, self-verification hooks, and mandatory sandboxing (harness disables itself without it).
Benchmark Highlights (company-reported)
| Benchmark | Computer | Pi | Hermes |
|---|---|---|---|
| Local Knowledge Work Bench | **82.6%** | 77.6% | 74.0% |
| BrowseComp (web research) | **66.7%** | 50.2% | 43.9% |
| Multimodal doc understanding | **65.1%** | 13.9% | 34.6% |
With post-trained PPLX 27B: **85.4%**. Also 51% less wall time and 70% fewer tokens than Pi.
The Economics
Chat was bursty; agents run for hours. Cloud metering makes that expensive — locally, marginal token cost approaches zero.
**Hybrid escalation math (Terminal Bench 2.1):**
- Fully local Qwen: **59.6%** (~$0)
- Local + Claude Opus 5 advisor: **73.0%** ($0.41/task) — recovers ~60% of the gap at ~⅔ the cost
- Frontier only: **82.4%** ($0.65/task)
Positioning vs. Ollama / DIY Stacks
Ollama solves inference; Portable Computer focuses on the **agent harness layer** and end-to-end integration (vLLM underneath, BYO-endpoint advanced mode). Nvidia frames it as fixing "the deeper you go, the deeper it gets" problem of DIY agentic setups.
Strategic Context
- **For Perplexity:** product economics independent of token metering; wedge into privacy-sensitive enterprises (law, healthcare, finance) amid ongoing publisher litigation
- **For Nvidia:** killer app for DGX Spark; Sparks scale via clustering (2 = frontier-class open models like DeepSeek; 4 = GLM 5.2; up to 8 demonstrated)
- Extends their June 2025 sovereign-AI partnership — same philosophy at desk scale: *intelligence you control, on hardware you own*
Open Questions
- Top benchmarks come from Perplexity's own internal eval (planned open-source)
- Compact models still trail frontier on hard reasoning
- Linux-only at launch; 24GB VRAM excludes most consumer PCs; no Apple silicon roadmap ("very focused right now on Nvidia hardware")
**Bottom line:** A credible push to move real agent workloads off the cloud — symbolized by a credit counter parked motionless at zero while an agent churns through tax documents on a desktop.