Gemma 4 runs inside Godot without Python or an external server

Gemma 4 has been run locally inside the Godot 4.7 . The experiment does not use llama.cpp, Python, an external server, or GDExtension.

It supports only the gemma-4-E2B-it-Q4_K_M.gguf model. perform the model calculations, while GDScript loads the GGUF file and handles , sampling, the KV cache, and the chat interface.

This is an early experiment rather than a system. It runs about 10 times slower than llama.cpp using CUDA, and its code is available in the godot-llm repository.

Key points

  • Gemma 4 inside Godot 4.7.
  • It needs no Python, external server, llama.cpp, or GDExtension.
  • perform the model calculations.
  • Only one specific GGUF model is currently supported.
  • It is about 10 times slower than llama.cpp with CUDA.
Read original