community pool / model answer

gemma-4-E4B-it

chat · google/gemma-4-E4B-it

57.3 tok/s · model median · 29 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
AMD Ryzen 7 8745H w/ Radeon 780M	4-bit	25.4 tok/s	reported	883	3.5 GB	32,768	llama.cpp
GTX 1060 3GB	4-bit	24.7 tok/s	reported	18,342	- GB	2,048	llama.cpp
Radeon 890M 32GB	4-bit	26.7 tok/s	reported	-	- GB	4,096	llama.cpp
Radeon AI Pro R9700 32GB	4-bit	137 tok/s	reported	-	- GB	2,048	llama.cpp
Radeon AI Pro R9700 32GB	8-bit	96 tok/s	reported	-	- GB	2,048	llama.cpp
Radeon AI Pro R9700 57GB	4-bit	25.3 tok/s	reported	982	- GB	2,048	llama.cpp
RTX 2080 8GB	4-bit	29 tok/s	reported	-	9.6 GB	2,048	ollama
RTX 3060 12GB	4-bit	73.3 tok/s	reported	207	7.3 GB	4,096	llama.cpp, ollama
RTX 3080 12GB	4-bit	130 tok/s	measured	-	9.8 GB	2,048	ollama
RTX3060+RTX2080-pooled-20GB 20GB	4-bit	76 tok/s	measured	-	9.6 GB	2,048	ollama
RX 570 4GB ×2	2-bit	22.3 tok/s	reported	280	- GB	2,048	llama.cpp
RX 570 4GB ×2	4-bit	28.7 tok/s	measured	254	- GB	2,048	llama.cpp
RX 9060 XT 16GB	8-bit	52.6 tok/s	reported	2,719	5.6 GB	8,192	llama.cpp
Ryzen AI Max 395 128GB	4-bit	57.3 tok/s	measured	-	10.5 GB	32,768	llama.cpp

Context tested

The table above is the tested context for each exact rig and quantization cell. Empty fields remain empty rather than being filled with an estimate.