Benchmark result
gemma-4-26B-A4B-it-qat-UD-Q4_K_XL on Radeon Instinct MI60 — 64.9 tok/s
Measured with llama.cpp b10384-6b517e82a on September 4, 2026.
What the model built
Open full screen →The coding scenario asks for a playable game in a single HTML file. This is exactly what the model returned, unedited.
How this run compares
2nd fastest of 2 runs of this model on Radeon Instinct MI60 · median 70.8 tok/s.
sort
- 176.8tok/s69.4llama.cpp · Q4_K
- 264.9tok/s64.7llama.cpp · Q4_Kthis run
Reproduce this run
toolllama.cpp b10384-6b517e82a
modelgemma-4-26B-A4B-it-qat-UD-Q4_K_XL (Q4_K)
context65,536
thinkingoff
samplingtemperature 0.5 · top_p 0.949999988079071 · top_k 64 · min_p 0.05000000074505806 · presence_penalty 0 · repeat_penalty 1
clientv0.4.54+97
Set those in llama.cpp, then:
llm-benchmark benchmark --model "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL" --tool "llama.cpp"The client prompts for context length and thinking mode, and for the KV cache dtype on Unsloth Studio. Sampling is left at the model default — the values above are what the tool reported using, not overrides the client sent.