98.4% of fine-tuned models use LoRA. That's not because it's best.

If you’ve fine-tuned an open-source model recently, you almost certainly used LoRA. And you’re not alone. An analysis of over 20,000 model cards on the Hugging Face Hub reveals a staggering statistic: 98.4% of checkpoints that use a parameter-efficient fine-tuning (PEFT) method use LoRA. It’s not just a narrow lead; it’s a near-total monopoly. The team at Hugging Face found a similar pattern on external image generation sites, where LoRA gobbled up 95% of identified PEFT checkpoints, leaving scraps for alternatives like LoCon and DoRA. Even scraping GitHub for PEFT import statements shows LoRA at 71.3%, with the next closest contender, LoHa, lagging at a paltry 3.7%.

This dominance feels less like a consensus on technical merit and more like a textbook case of a feedback loop. LoRA was an early, genuinely effective method. That initial success spawned tutorials, documentation, and deep integration into downstream tools, which in turn made it everyone’s default choice. The Hugging Face team poses the uncomfortable question this reality demands: are we all leaving performance on the table? It’s a valid fear, especially when the PEFT library itself now contains over 40 distinct techniques, and a parade of research papers claim their new method dethrones LoRA on standard benchmarks.

But trusting those academic claims is its own trap. The blog post points to a classic research pathology: the intense pressure to publish state-of-the-art numbers can bias experiments, whether consciously or not. A researcher might spend weeks tuning their novel technique while giving LoRA a cursory, suboptimal learning rate sweep. One study even demonstrated that simply tuning the learning rate allowed LoRA to match supposedly superior methods. The comparison problem is also a mess of incompatibility, with each paper picking different baselines, benchmarks, and often not releasing easy-to-use code.

So what’s the solution? The answer isn’t to blindly jump to the newest method with the flashiest chart, but to get systematic. The post sets up the argument that the real bottleneck isn’t a lack of LoRA alternatives—it’s a lack of reliable, accessible tools to compare them. It’s not about whether a technique can beat LoRA in one contrived setting; it’s about whether it beats LoRA on your data, with your resources, without a PhD student spending a month finessing the hyperparameters. That’s a much harder, more boring problem to solve, but arguably the one that actually matters.