量子位 (微信公众号)T2·78 pts
- Karpo Decision 31B targets high-frequency, low-latency agent decisions, runs on a single RTX PRO 6000, and reports a 94 ms median response on public requests
- karpo-browser uses a System 1 + System 2 architecture with a 27B Act action model and the KBB web-action evaluation set
- On Browserbase Benchmark v2 it reports 89% success, 201 seconds per task, and $0.096 model cost per task, well ahead of the Claude Fable 5.2 setup