Perplexity's Portable Computer: A Local-First Agent on Qwen3.8 27B, Near-Zero Cost
机器之心 · wechat · 2026-08-30
Perplexity's Portable Computer is an agent that runs fully local by default—model, harness, conversations and trajectories stay on-device, sensitive data never leaves, and local inference is free; web search, connectors, or cloud advisor upgrades only cross the device boundary with user approval. The key idea is co-designing model and framework: instead of making a small local model (Qwen3.8 27B) drive a harness built for frontier models, the framework is tailored to it—modular skills loaded on demand (performance degrades past 100k tokens), common MCP servers converted to compact CLI tools, self-verification hooks, and mandatory OS-level sandboxing.
On an NVIDIA DGX Spark with Qwen3.8 27B, Portable Computer beats open-source harnesses on BrowseComp web research (66.7% vs Hermes 50.2% / Pi 43.9%) and ParseBench-100 document understanding (65.1%) with the lowest runtime and token usage. A frontier-model "advisor" upgrade lifts TerminalBench 2.1 from 59.6% to 73.0% at $0.415 per deployment (raw Claude Opus 5: 82.4% at $0.65). Two-stage post-training on synthetic RL environments produced PPLX 27B, raising the Local Knowledge WorkBench score from 82.6% to 85.4%.
More from coding & agent
- Simon Willison: code mode is the future of tool calling, if it can be sandboxed — springrod · 2026-10-16
- Agent named 'Ilya' reportedly scores 1.1pt higher on benchmarks — daniel_mac8 · 2026-10-16
- Gemini CLI fix: substring matching inlined unrequested binary files into model context — priyanshu1976 · 2026-10-16
- Building AI agents on real healthcare claims data: the domain work that makes them work — alex_verem · 2026-10-16
- OpenAI stealth-testing mysterious internal model 'Aeon' in Codex desktop app and Dots — cedric_chee · 2026-10-16
- Vibe-coded iPhone Duo sniping bot built with Codex in 17 minutes, 41 tests passing — xiaohu · 2026-10-16