Hugging Face上的国产模型越来越多，但这事没那么简单 (EN)

Lately when I browse Hugging Face, I’ve noticed a clear increase in recommended Chinese models. Kimi, Qwen series, MiniMax, DeepSeek V4 preview… the list keeps growing.

Is this a good thing? Partly.

More models means more choices, which is good for developers. But I’ve been burned by several “looks strong, actually painful to use” models recently, so I want to share some honest thoughts.

The first issue is benchmark dataset overfitting. Some domestic models score very high on specific benchmarks but perform mediocrely in practice—because their training data was deliberately similar to the evaluation sets. Hugging Face download counts and actual user reputation often diverge significantly.

The second issue is “open source in name only.” Some models claim to be open source but only release inference code, keeping training data, training scripts, and model details under lock. This kind of “open source” actually has limited value to the community.

The third issue is relatively minor but still annoying: documentation and example code quality varies widely. Some model READMEs are detailed and thorough; others can’t even pass basic English machine translation checks.

My suggestion: don’t just look at downloads and rankings. Check developer forums (like Hugging Face discussions) for genuine user feedback. Truly good models often have lagging reputations—it takes time for a consensus to form once enough people have tried them and bumped into issues.

Among domestic models, there are indeed some genuinely well-done releases, like the Qwen series and DeepSeek. But run small-scale tests before going to production.

Tags： AI开源生态、Hugging Face、Qwen、DeepSeek、模型评测