Key Info

Hugging Face's Julien Chaumond shared that Pi now offers a "Virtual Models" feature that automatically routes each query to a local Qwen 3.8 27B model (via Llama.app) or to remote Kimi K3 / GLM 5.3, based on query complexity and current machine load.

Highlights

  • Routing decision depends on two factors: query complexity and the local machine's current load
  • Local inference runs Qwen 3.8 27B inside Llama.app
  • Remote options include Kimi K3 and GLM 5.3
  • Removes the need for manual model switching
  • Credits given to badlogicgames, mitsuhiko and the ggml team