Details
- Perplexity announces Portable Computer, a local-first agent designed for private and cost-effective knowledge work on user hardware.
- The system runs an on-device 27B parameter model that scores 82.6% on Perplexity’s real knowledge work harness, outperforming open-source harnesses Pi and Hermes.
- A post-trained variant, PPLX 27B, reaches 85.4% on the same benchmark, reflecting tuning inside the Computer harness on synthetic tasks derived from real usage patterns (without real user data).
- Portable’s architecture combines a minimal system prompt, skills that load on demand, compact CLI-based connectors instead of MCP servers, self-verification, and an always-on sandbox for safer tool use.
- On 1,266 BrowseComp web-research tasks, Computer achieves 66.7% accuracy versus 50.2% for Pi and 43.9% for Hermes on their respective search providers, while using the least wall time and fewest tokens.
- Inference and access to private documents remain local by default; only search queries touch the web, emphasizing privacy and reduced cloud dependence.
- Parsing of documents runs entirely on-device, so sensitive files never leave the machine during understanding or extraction tasks.
- On ParseBench-100, Computer scores 65.1% compared with 34.6% for Hermes and 13.9% for Pi, again with the lowest time and token usage.
- The local model can selectively escalate to a frontier advisor under user control, with personally identifiable information flagged and only text guidance sent to the cloud.
- On Terminal Bench 2.1, escalation from fully local to frontier-assisted lifts the score from 59.6% to 73.0% at approximately $0.415 per rollout, recovering about three-fifths of the performance gap at roughly two-thirds of the frontier-only cost.
- Perplexity highlights that continued advances in models, chips, and end-user devices will further expand what local-first agents can do, and commits to opening its benchmarks soon.
- The full technical and benchmarking details are provided in a linked Perplexity post about Portable Computer and PPLX 27B.
Impact
Perplexity’s Portable Computer pushes the emerging local-first agent model by showing that a tuned 27B on-device system can approach frontier-agent performance while preserving privacy and lowering per-task costs. By pairing a purpose-trained model with a harness optimized for smaller LLMs and optional, user-gated escalation to cloud advisors, Perplexity moves closer to a hybrid standard where most agentic work runs locally and only the hardest problems tap expensive frontier models. This approach could pressure other agent platforms to offer similar local-first options and clearer cost–performance tradeoffs.
