Your line that a model topping one leaderboard can look mediocre on another matches my experience: a coding agent I picked off a leaderboard flopped on our repo. I make the final call by replaying our own merged work, as I wrote up here: kartiknvjk.hashnode.dev/i-picked-a-coding-agent-o…. Do you have numbers for the 70B model's quality loss on that 4GB card?