Perplexity Introduces Portable Computer Local AI Agent
Perplexity shares benchmarks for its on-device 27B model harness.
Perplexity posted that Portable Computer is a local-first agent for private and cost-effective work. With an on-device 27B model, its harness scores 82.6% on real knowledge work, beating open-source harnesses Pi and Hermes, the company said. The company said its post-trained PPLX 27B reaches 85.4%. Aravind Srinivas noted advisor escalation to a cloud frontier model lifts Terminal Bench 2.1 scores from 59.6% to 73.0% at $0.415 per rollout. The posts highlight local parsing and inference for privacy with optional cloud help.
Combined views
33.7K
3 posts, first seen 6h ago
Perplexity Introduces Portable Computer Local AI Agent
Perplexity shares benchmarks for its on-device 27B model harness.
Perplexity posted that Portable Computer is a local-first agent for private and cost-effective work. With an on-device 27B model, its harness scores 82.6% on real knowledge work, beating open-source harnesses Pi and Hermes, the company said. The company said its post-trained PPLX 27B reaches 85.4%. Aravind Srinivas noted advisor escalation to a cloud frontier model lifts Terminal Bench 2.1 scores from 59.6% to 73.0% at $0.415 per rollout. The posts highlight local parsing and inference for privacy with optional cloud help.
Reactions
3 postsAS 
Aravind Srinivas@AravSrinivasPerplexity's harness for Portable Computer leads to superior results on benchmarks like BrowseComp compared to other local harnesses.
- likes: 17
- replies: 1
- bookmarks: 0
- reposts: 0
AS Aravind Srinivas@AravSrinivasLocal models are still quite far from the frontier for complete zero dependence on the frontier. The advisor escalation to a cloud-based frontier model approach addresses this issue, allowing for a hybrid agentic inference system. On Terminal Bench 2.1, escalation lifts the…
- likes: 8
- replies: 1
- bookmarks: 0
- reposts: 0
PE Perplexity@perplexity_aiNew research: Portable Computer is a local-first agent for private and cost-effective work. With an on-device 27B model, our harness scores 82.6% on real knowledge work, beating open-source harnesses Pi and Hermes. Our post-trained PPLX 27B reaches 85.4%.
- likes: 302
- replies: 23
- bookmarks: 74
- reposts: 23
