For my recent LLM leaderboard, I sum up all time total downloads (or likes) across model variants, and then divide it by the age of that model. I.e. "time decay" for popularity This gives a more time-agnostic metric for the popularity of that model. In an ideal ranking, older models that are not popular anymore should be demoted, like 2 year old Llama 3 models. If you don't do that, they might still occupy top 10 needlessly, despite having been replaced by e.g. qwen in practice Thanks to that, qwen-3-6b which came up 1 year ago and has 150m downloads can surpass llama-3-1-8b which came up 2 years ago and has 200m downloads More notes on my post: