The GemmaPro model is based on Gemma3 4B, using Q4 quantization technology. After weight reorganization, the model size is only about 2.5 GB, requiring just 4 GB RAM to run smoothly. This means most office computers or high-end phones have the potential to use it.

Any GPU with 4GB or more VRAM can fully load the model. On computers without a GPU, the model can achieve 7-8 tokens per second in inference speed, which is close to normal human speaking speed, making conversations feel smooth without noticeable delay. When using a GPU, the speed is even faster.