Benchmark result
Qwen3.8-27B-UD-Q3_K_XL on NVIDIA GeForce RTX 5070 Ti — 52.7 tok/s
Measured with llama.cpp b11469-3cb90bf49 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; the frame only adds a line that tells this page how tall it is.
How this run compares
15th fastest of 18 runs of this model on NVIDIA GeForce RTX 5070 Ti · median 67.7 tok/s.
sort
- 178.4tok/s68.1llama.cpp · Q3_K
- 275.1tok/s88.6llama.cpp · Q3_K
- 374.5tok/s63.3llama.cpp · Q3_K
- 471.9tok/s86.6llama.cpp · Q3_K
- 570.3tok/s86.3llama.cpp · Q3_K
- 668.4tok/s86.1llama.cpp · Q3_K
- 768.1tok/s66.2llama.cpp · Q3_K
- 868.1tok/s66.1llama.cpp · Q3_K
- 968.0tok/s88.1llama.cpp · Q3_K
- 1067.4tok/s79.5llama.cpp · Q3_K
- 1166.9tok/s87.0llama.cpp · Q3_K
- 1266.2tok/s86.6llama.cpp · Q3_K
- 1366.2tok/s86.1llama.cpp · Q3_K
- 1465.9tok/s75.7llama.cpp · Q3_K
- 1552.7tok/s85.7llama.cpp · Q3_Kthis run
- 1646.5tok/s45.1llama.cpp · Q3_K
- 1744.7tok/s85.4llama.cpp · Q3_K
- 1843.1tok/s62.0llama.cpp · Q3_K
Reproduce this run
toolllama.cpp b11469-3cb90bf49
modelQwen3.8-27B-UD-Q3_K_XL (Q3_K)
context65,536
samplingtemperature 1 · top_p 0.949999988079071 · top_k 20 · min_p 0 · presence_penalty 0 · repeat_penalty 1
clientv0.4.54+97
Set those in llama.cpp, then:
llm-benchmark benchmark --model "Qwen3.8-27B-UD-Q3_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.