Benchmark result
Qwen3.8-27B-Q4_K_XL on AMD Radeon RX 7900 XTX — 72.8 tok/s
Measured with llama.cpp b182-c1d0e7a on August 29, 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; the frame only adds a line that tells this page how tall it is.
How this run compares
11th fastest of 15 runs of this model on AMD Radeon RX 7900 XTX · median 73.3 tok/s.
sort
- 177.0tok/s74.9llama.cpp · Q4_K
- 275.6tok/s82.9llama.cpp · Q4_K
- 375.6tok/s82.8llama.cpp · Q4_K
- 475.3tok/s86.5llama.cpp · Q4_K
- 574.4tok/s74.2llama.cpp · Q4_K
- 674.3tok/s84.7llama.cpp · Q4_K
- 773.8tok/s87.1llama.cpp · Q4_K
- 873.3tok/s79.1llama.cpp · Q4_K
- 973.3tok/s82.0llama.cpp · Q4_K
- 1073.0tok/s83.8llama.cpp · Q4_K
- 1172.8tok/s82.9llama.cpp · Q4_Kthis run
- 1272.2tok/s67.8llama.cpp · Q4_K
- 1371.3tok/s85.9llama.cpp · Q4_K
- 1470.2tok/s82.9llama.cpp · Q4_K
- 1554.1tok/s84.6llama.cpp · Q4_K
Reproduce this run
toolllama.cpp b182-c1d0e7a
modelQwen3.8-27B-Q4_K_XL (Q4_K)
context65,536
thinkingon
clientv0.4.44+97
Set those in llama.cpp, then:
llm-benchmark benchmark --model "Qwen3.8-27B-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.