Cactus Compute's new model packs tool calling into 45M parameters and a 14MB binary.
Liquid AI's 2.6B model plans, calls tools, and runs fully on-device with a 128K context.