Benchmark result
Qwen3.8-27B-UD-IQ3_S on NVIDIA GeForce RTX 4080 SUPER — 76.9 tok/s
Measured with llama.cpp b10729-458681e1d on September 5, 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
39th fastest of 48 runs of this model on NVIDIA GeForce RTX 4080 SUPER · median 79.1 tok/s.
sort
- 183.4tok/s86.1llama.cpp
- 283.2tok/s85.0llama.cpp
- 383.1tok/s83.8llama.cpp
- 482.8tok/s83.2llama.cpp
- 582.7tok/s81.9llama.cpp
- 682.3tok/s84.0llama.cpp
- 782.3tok/s83.0llama.cpp
- 881.6tok/s84.3llama.cpp
- 981.5tok/s85.5llama.cpp
- 1081.4tok/s84.7llama.cpp
- 1181.1tok/s80.0llama.cpp
- 1281.1tok/s83.8llama.cpp
- 1381.0tok/s87.5llama.cpp
- 1480.6tok/s83.9llama.cpp
- 1580.5tok/s83.4llama.cpp
- 1680.3tok/s82.4llama.cpp
- 1780.3tok/s81.6llama.cpp
- 1880.0tok/s82.9llama.cpp
- 1979.8tok/s84.1llama.cpp
- 2079.7tok/s81.2llama.cpp
- 2179.3tok/s82.6llama.cpp
- 2279.2tok/s83.0llama.cpp
- 2379.1tok/s84.1llama.cpp
- 2479.1tok/s88.3llama.cpp
- 2579.1tok/s83.6llama.cpp
- 2679.0tok/s82.6llama.cpp
- 2779.0tok/s66.6llama.cpp
- 2878.9tok/s84.5llama.cpp
- 2978.7tok/s85.2llama.cpp
- 3078.6tok/s79.7llama.cpp
- 3178.1tok/s79.4llama.cpp
- 3277.7tok/s83.3llama.cpp
- 3377.6tok/s84.9llama.cpp
- 3477.3tok/s86.4llama.cpp
- 3577.3tok/s84.3llama.cpp
- 3677.1tok/s82.5llama.cpp
- 3777.1tok/s86.5llama.cpp
- 3876.9tok/s81.3llama.cpp
- 3976.9tok/s86.1llama.cppthis run
- 4076.7tok/s83.6llama.cpp
- 4176.7tok/s85.1llama.cpp
- 4274.6tok/s84.8llama.cpp
- 4374.2tok/s82.3llama.cpp
- 4474.1tok/s80.2llama.cpp
- 4573.9tok/s80.5llama.cpp
- 4673.5tok/s84.5llama.cpp
- 4773.3tok/s85.3llama.cpp
- 4872.3tok/s81.2llama.cpp
Qwen3.8-27B-UD-IQ3_S on other hardware
Reproduce this run
toolllama.cpp b10729-458681e1d
modelQwen3.8-27B-UD-IQ3_S
context65,536
samplingtemperature 0.699999988079071 · top_p 0.800000011920929 · top_k 20 · min_p 0 · presence_penalty 1.5 · repeat_penalty 1
clientv0.4.55+97
Set those in llama.cpp, then:
llm-benchmark benchmark --model "Qwen3.8-27B-UD-IQ3_S" --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.