Benchmark result
nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16:MXFP4_F16 on NVIDIA GeForce RTX 3080 Ti +1 more — 12.4 tok/s
Measured with llama.cpp manual on October 11, 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.
Reproduce this run
toolllama.cpp manual
modelnvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16:MXFP4_F16
context65,536
thinkingon
think limitnone
samplingtemperature 0.6000000238418579 · top_p 0.949999988079071 · top_k 40 · min_p 0.05000000074505806 · presence_penalty 0 · repeat_penalty 1
clientv0.4.91+102
Set those in llama.cpp, then:
llm-benchmark benchmark --model "nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16:MXFP4_F16" --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.