Analysis shows repacked Gemma 4 E2B model outperforms Google's 4-bit export

A technical analysis compares two 4-bit versions of Google's Gemma 4 E2B language model. The test was conducted on a Colab TPU v5e runtime using a pure-JAX engine. The repackaged model, which retains the original trained quantization grid, demonstrated significantly higher accuracy in next-token predictions compared to Google's official export. Google's export applies a second rounding step to the weights, diverging from the model's training. The repackaged version also matched original model outputs more closely while requiring 0.8 GB less storage.