Gemma 4 26B A4B

MXFP4 running on RX 6700 XT

PERFORMANCE / TOKENS PER SECONDRDNA 2

Prompt processing552tok/s

#9 in index

Token generation22tok/s

#20 in index

Run fingerprint

Enough context to make the number useful.

ModelGemma 4 26B A4BMoE architecture

QuantizationMXFP4Exact submitted variant

GPURX 6700 XT12 GB VRAM

GPU architectureRDNA 2Metal backend

Test context512prompt tokens

Model familyMoECPU expert offload

Prompt processing and generation are reported separately so memory bandwidth and decode behavior stay visible.

Quantization, GPU family, VRAM and context travel with the result instead of being hidden behind one score.

This run was reproduced by ToshLLM Lab.