community pool / model answer

gemma-4-26B-A4B-it

chat · google/gemma-4-26B-A4B-it

60.6 tok/s · model median · 43 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
Arc Pro B70 32GB	8-bit	43.3 tok/s	reported	393	- GB	2,048	llama.cpp
AMD Ryzen 7 8745H w/ Radeon 780M	4-bit	30.6 tok/s	reported	1,534	4 GB	256,000	llama.cpp
M1 Max 64GB	8-bit	49 tok/s	reported	462	- GB	2,048	llama.cpp, ollama
M3 Ultra 512GB	4-bit	76.9 tok/s	reported	279	- GB	2,048	mlx
Radeon AI Pro R9700 32GB ×2	4-bit	87.3 tok/s	reported	375	- GB	794	llama.cpp
Radeon AI Pro R9700 32GB ×2	8-bit	99.3 tok/s	measured	-	- GB	262,144	llama.cpp
Radeon AI Pro R9700 32GB ×3	8-bit	73.7 tok/s	measured	298	35.4 GB	794	llama.cpp
RTX 3090 24GB ×2	8-bit	154.1 tok/s	reported	42	45.3 GB	262,144	vllm
RTX 4070 12GB	4-bit	20.4 tok/s	measured	1,661	- GB	2,048	vllm
RTX 5090 32GB	8-bit	282 tok/s	measured	30	- GB	2,048	vllm
RTX PRO 6000 Blackwell 96GB	8-bit	276.7 tok/s	reported	28	- GB	8,192	vllm
RX 570 4GB ×2	4-bit	12.5 tok/s	measured	760	- GB	2,048	llama.cpp
Ryzen AI Max 395 128GB	4-bit	60.7 tok/s	measured	-	21.5 GB	32,768	llama.cpp
Ryzen AI Max 395 128GB	6-bit	55.5 tok/s	reported	-	23.8 GB	16,384	llama.cpp
Ryzen AI Max 395 128GB	8-bit	46.1 tok/s	reported	746	27.8 GB	16,384	llama.cpp
Tesla P100-PCIE-16GB 16GB ×2	5-bit	54.8 tok/s	reported	-	- GB	2,048	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.