35B Benchmark on RTX 5060 Ti

On a combination of AMD 9700X, RTX 5060 Ti 16GB, and 32GB RAM, the author built llama.cpp from source and benchmarked Qwen3.6-35B-A3B-UD-Q5_K_M.

• Command used: llama-bench.exe -m models\\Qwen3.6-35B-A3B-UD-Q5_K_M.gguf -ngl 99 --n-cpu-moe 22 -d 131072 -p 512 -n 128 --cache-type-k q8_0 --cache-type-v q8_0 -fa 1 -mmp 0
• Model file size: 24.63 GiB
• Parameter count: 34.66B
• GPU detection info: 1 CUDA device, VRAM 16,310 MiB, compute capability 12.0

The benchmark results were as follows.

• prompt processing: 628.10 t/s (pp512 @ d131072)
• token generation: 32.56 t/s (tg128 @ d131072)

The author reported achieving about a 10% performance improvement over the prebuilt llama.cpp.