Bonsai models are pure hype: Bonsai-8B is MUCH dumber than Gemma-4-E2B
I'm using the https://github.com/PrismML-Eng/llama.cpp fork for Bonsai, regular llama.cpp for Gemma. Without embedding parameters: Gemma 4 has 2.3B at 4.8 bpw (Q4_K_M) = 1104 MB Bonsai-8B has 6.95B at 1.125 bpw (Q1_0) = 782 MB (-29% smaller) …