A Reddit user has proposed a practical alternative: instead of repeatedly uploading large numbers of “uncensored” versions for every base model, publish portable LoRA adapters. The post says Hugging Face already hosts several million such models, meaning duplicate storage could create substantial waste that the community should address.
What this is
LoRA (Low-Rank Adaptation) is a type of small parameter patch that can be attached to base models with the same architecture to alter their output style or behavior without retraining the entire model. The author argues that a single adapter can work with different quantizations—for example, multiple quantized versions of Qwen3.6-35B-A3B—and proposes creating a categorized LoRA repository with comparable models.
This is not a formal product plan announced by Hugging Face; it is a community proposal initiated by a Reddit user. It highlights storage, reuse, and distribution-efficiency issues in the open-model ecosystem.
Industry view
The case for this approach is straightforward: adapters are usually far smaller than full models, while repeatedly uploading complete weights consumes space and increases download and retrieval costs. If adapters can be reused across quantized versions, maintainers can publish once and users can experiment across more model configurations, moving the open-model ecosystem closer to modularity.
The risks are equally clear. First, whether an adapter is universal depends on the base model’s architecture and training lineage; not every model can use it directly. Second, “uncensored” adapters may bypass safety restrictions and amplify harmful-content risks, so platforms need to assess carefully whether they are willing to host them long term. Third, compatibility between adapters and base models, version mismatches, and source credibility will all increase the cost of filtering options for users. Without standardized testing, fewer repositories do not automatically mean lower risk.
Impact on regular people
For enterprise IT: If the approach matures, organizations can reduce the space devoted to duplicate model weights and more easily test different model behaviors on their own servers. Before deployment, however, they still need processes for reviewing provenance, licensing, and security.
For individual professionals: For employees who need to run models locally, adapters may lower the cost of switching between models and quantized versions. For product and operations staff, understanding how base models and adapters work together will gradually become a basic skill.
For the consumer market: More efficient storage and distribution of open models could indirectly drive cheaper local AI services. But the proliferation of uncensored models could also increase pressure around content governance. Platforms and product providers cannot pursue usability alone.