Comparing a 7B model against a 70B model in a benchmark table has always been a bit pointless. The 70B wins almost every time - but if you're running locally on a consumer GPU, that comparison tells you nothing useful about what's actually available to you.
HuggingFace has fixed this with a model size filter on its benchmark datasets. You can now narrow benchmark results by parameter count - the number of numerical weights in a model, which determines both how capable it tends to be and how much memory it needs to run. A 7B model (7 billion parameters) typically fits on a single consumer GPU. A 70B model generally requires either enterprise hardware or some form of quantization (a compression technique that trades a small amount of accuracy for much lower memory use).
The update applies across HuggingFace's Open LLM Leaderboard and related benchmark collections, which track performance on tests like MMLU (a knowledge and reasoning test spanning 57 academic subjects) and HumanEval (a benchmark for coding ability measured by passing automated test suites).
For developers and researchers evaluating which local model to deploy on a given machine, the filter closes a real usability gap. The question "which model performs best at my hardware tier" now has a direct answer without manually cross-referencing parameter counts against scores. Small change, practical payoff.