1/ Introducing Orvena 4B: a 4B model for on-device tool calling, trained on calendar, alarms, places, web and MCP tasks in 30 languages. On our 547-task bench it passes 80.4%, against 63.1% for Qwen3.5-4B and 19.0% for Apple's on-device model. Open weights, Apache 2.0. https://t.co/m0TcSPltbc
Nov 5, 2026
Views56.3k
Comments2
Reposts0