artificialanalysis.ai for benchmarks

from huggingface
Qwen3.6 35B A3B (It's good enough for a quick script)
Gemma4 26B A4B
(maybe + heretic)

Ollama does not show you default quantization+params. I found it too opaque in general.
I personally use llama-cpp
with --models-preset /path/to/a-config.ini