Benchmark result
Cyber-Tiel-Coder-35B-A3B-MTP-UD-IQ3_XXS on Tesla P100-PCIE-16GB — 50.2 tok/s
Measured with llama.cpp b1-5266f24 on September 22, 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 41 runs of this model on Tesla P100-PCIE-16GB · median 59.2 tok/s.
sort
- 168.5tok/s72.4llama.cpp
- 267.3tok/s85.9llama.cpp
- 367.2tok/s85.1llama.cpp
- 466.5tok/s84.2llama.cpp
- 566.4tok/s84.7llama.cpp
- 666.1tok/s82.8llama.cpp
- 766.1tok/s85.0llama.cpp
- 866.1tok/s84.4llama.cpp
- 965.9tok/s83.8llama.cpp
- 1065.9tok/s84.2llama.cpp
- 1165.6tok/s83.5llama.cpp
- 1265.5tok/s83.6llama.cpp
- 1365.0tok/s80.9llama.cpp
- 1464.9tok/s66.3llama.cpp
- 1563.3tok/s83.7llama.cpp
- 1663.2tok/s85.6llama.cpp
- 1763.0tok/s84.8llama.cpp
- 1860.8tok/s85.7llama.cpp
- 1960.5tok/s81.7llama.cpp
- 2060.4tok/s81.6llama.cpp
- 2159.2tok/s86.6llama.cpp
- 2258.6tok/s79.6llama.cpp
- 2358.5tok/s86.5llama.cpp
- 2458.3tok/s87.3llama.cpp
- 2558.1tok/s75.8llama.cpp
- 2658.0tok/s69.3llama.cpp
- 2757.8tok/s77.9llama.cpp
- 2857.5tok/s83.2llama.cpp
- 2957.4tok/s78.5llama.cpp
- 3057.3tok/s83.4llama.cpp
- 3157.0tok/s83.6llama.cpp
- 3256.8tok/s77.8llama.cpp
- 3356.6tok/s77.3llama.cpp
- 3456.0tok/s81.9llama.cpp
- 3556.0tok/s78.3llama.cpp
- 3655.5tok/s76.7llama.cpp
- 3755.1tok/s80.7llama.cpp
- 3851.3tok/s69.7llama.cpp
- 3950.2tok/s83.7llama.cppthis run
- 4048.6tok/s74.0llama.cpp
- 4147.3tok/s76.1llama.cpp
Cyber-Tiel-Coder-35B-A3B-MTP-UD-IQ3_XXS on other hardware
Reproduce this run
toolllama.cpp b1-5266f24
modelCyber-Tiel-Coder-35B-A3B-MTP-UD-IQ3_XXS
context131,072
thinkingon
samplingtemperature 0.6000000238418579 · top_p 0.949999988079071 · top_k 20 · min_p 0 · presence_penalty 0 · repeat_penalty 1
clientv0.4.80+98
Set those in llama.cpp, then:
llm-benchmark benchmark --model "Cyber-Tiel-Coder-35B-A3B-MTP-UD-IQ3_XXS" --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.