CADA quantization more than doubles Gemma 4 E4B's reasoning score

A quantization method called CADA pushed Gemma 4 E4B's reasoning score from 28.9% to 69.5% at the same size, per its Hugging Face model card. That's a 40.6-point jump at the same 3.32 GiB budget, the range engineers target to fit language models on phones and laptops without a GPU.

Standard 2-bit quantization (IQ2_XXS) squeezes every tensor to the same precision. CADA instead searches for which tensors can tolerate 2-bit compression and which need more bits. It spends the same total byte budget unevenly across the model, per the model card.

The 3.32 GiB result is a 76.32% cut from the 14.02 GiB BF16 source model, the card says. It doesn't say how the reasoning benchmark was scored, or whether anyone besides ByteOtter has run the comparison.

For anyone shipping small models on-device, that's the specific number worth watching for independent replication before trusting a 2-bit checkpoint on reasoning-heavy work.