Benchmark result
Qwen3.8-27B on NVIDIA GeForce RTX 3090 — 47.0 tok/s
Measured with llama.cpp b11900-74b104c9b on November 13, 2026.
Support llm-bench.io
llm-bench.io is free. If it helped you, I'd appreciate your support to cover the infrastructure costs.
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
8th fastest of 15 runs of this model on NVIDIA GeForce RTX 3090 · median 47.0 tok/s.
sort
- 194.9tok/s22.7vLLM
- 256.6tok/s88.2llama.cpp
- 354.0tok/s90.7Strata
- 453.5tok/s82.3llama.cpp
- 552.6tok/s79.7llama.cpp
- 650.9tok/s86.9llama.cpp
- 750.0tok/s88.3Strata
- 847.0tok/s86.4llama.cppthis run
- 938.4tok/s90.9Strata
- 1037.7tok/s88.0Strata
- 1137.3tok/s88.2llama.cpp
- 1236.9tok/s89.6llama.cpp
- 1332.4tok/s88.5llama.cpp
- 1432.1tok/s90.3llama.cpp
- 1516.5tok/s86.2LM Studio · Q4_K_M
Reproduce this run
toolllama.cpp b11900-74b104c9b
modelQwen3.8-27B
context120,000
thinkingon
effortxhigh
think limitnone
samplingtemperature 1 · top_p 0.949999988079071 · top_k 20 · min_p 0 · presence_penalty 0 · repeat_penalty 1
clientv0.4.98+102
Set those in llama.cpp, then:
llm-benchmark benchmark --model "Qwen3.8-27B" --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.