Something notable is that Qwen3.5:0.8B does better on benchmarks than GPT3.5. Runs much faster on local hardware than GPT3.5 on release. However Qwen3.5:0.8B dumber and slower than GPT3.5. It's dumber: it can do 3*3, but if asked to explain it in terms of the definition (i.e. 3+3+3=9) it fails. It's slower: It's a thinking model so your 900T/S are mainly spent "thinking" most of the time it will just repeat until it hangs.

It pretty obvious that this reasoning scaling is a mirage, parameters are all you need. Everything else is mostly just wasting time while hardware get better.