Hugging Face is revisiting one of the default choices in open-model development: whether LoRA remains the best fine-tuning approach for a given task.
LoRA became popular because it makes adaptation cheaper and easier, but the growing variety of model sizes, hardware setups, and deployment targets is pushing teams to test alternatives more carefully.
For practitioners, the useful takeaway is not that one method wins everywhere. It is that fine-tuning strategy should be benchmarked against the task, budget, and serving constraints rather than copied from convention.