from what I gather it's finetuned to use OpenHand specifically so shows value on thsoe benchmark that target a whole system as a blackbox (i.e. agent + llm) more than directly target the llm input/outputs

Yup the 1st comment says this https://www.reddit.com/r/LocalLLaMA/comments/1kryybf/mistral...