Back to home

Compare

Comparing: Hugging Face Audit: 14% of Open-Source AI Model Files Are Mislabeled & Hugging Face 审计结果:14% 的开源 AI 模型文件名实不符

AEN
Hugging FaceGGUFllama.cpp·

Hugging Face Audit: 14% of Open-Source AI Model Files Are Mislabeled

What this is

Last week, an independent developer published a Reddit audit of Hugging Face (the world's largest open-source model hosting platform): he sampled 443 GGUF model files intended for local inference across 25 repositories, and found 64 — roughly 14.4% — where "the filename says one thing, the actual precision says another."

What is GGUF? In short, it's a format that compresses and packages large models so personal computers can run them. "Quantization" can be thought of as a discount on model precision — the lower the level (e.g., 2-bit), the smaller the model and the lower the GPU requirement, but the more the model's comprehension ability takes a hit. The core finding of the audit: when certain tensor dimensions inside a model are not divisible by 256, llama.cpp (the most popular open-source inference engine) silently substitutes a roughly 4.5-bit version during compression, while the filename still reads 2-bit.

The most striking case is a MoE-architecture (Mixture of Experts — splitting a model into sub-modules that activate on demand) large model uploaded by multiple quantizers: four versions labeled 2.06 to 2.56 bit, but all four measured at 4.58 bit — the same file, wearing four different "low-VRAM" signs.

Industry view

The auditor's tone is restrained — he stresses this is a design behavior of llama.cpp dating back to 2023, and the quantization tool does print a warning to its log. The problem is that the warning lives in the quantization log, while the vast majority of users download pre-built GGUF files from others and never see it.

But there is pushback in the community. Some argue this isn't a "tool bug" but a "trust chain bug" — the precision stated in the model card (the model's documentation page), the precision in the filename, and the precision described in the metadata all three agree on pointing users to a version they can't actually obtain. When bartowski (one of the most trusted quantizers in the field) and other top players fall into the trap, the problem is no longer a single case of carelessness — it's the entire local AI ecosystem lacking file-level verification.

Another risk worth flagging: this audit covered only 25 repositories, while the total number of GGUF files on Hugging Face is in the tens of thousands. We have reason to believe the 14.4% figure is likely a significant underestimate — this is sampling bias, not a conclusion.

Impact on regular people

For enterprise IT: if your company is evaluating "cut costs by deploying open-source models locally," treat claims like "this 2-bit model only needs 16GB of VRAM" with verification from today onward — otherwise compute budgets may blow up.

For working professionals: practitioners running large models on MacBooks or consumer GPUs — programmers, consultants, research assistants — if you've recently found that "they said it would save VRAM, but it still won't run," now you know roughly why.

For the consumer market: vendor pitches for "AI laptops running 7-billion-parameter models locally" are flooding 2025, and the transparency problem around quantization precision will eventually reach consumers — if you buy one and it can't run, the fault lies not in the computer but in the inflated files upstream.

BZH
Hugging FaceGGUFllama.cpp·

Hugging Face 审计结果:14% 的开源 AI 模型文件名实不符

这是什么

上周,海外一位独立开发者在 Reddit 公开了一份针对 Hugging Face(全球最大的开源模型托管平台)的审计报告:他在 25 个仓库里抽查了 443 份供本地运行使用的 GGUF 模型文件,发现其中 64 份——约 14.4%——"文件名说一套,实际精度是另一套"。

什么是 GGUF?简单说就是把大模型"压缩打包"后让个人电脑能跑得起来的格式;"量化"(quantization)可以理解为给模型精度打折,等级越低(比如 2-bit)模型体积越小、对显卡要求越低,相应的理解能力也会打折扣。这次审计的核心发现是:当模型内部某些张量(tensor,可以理解为模型中的数据块)尺寸不能被 256 整除时,llama.cpp(最主流的开源推理引擎)在压缩时会悄悄用一个约 4.5-bit 的版本替代,但文件名依然写着 2-bit。

最典型的案例是一款被多个量化师分别上传的 MoE 架构(Mixture of Experts,混合专家——把模型拆成多个子模块按需调用)大模型:4 个标注为 2.06 到 2.56 bit 的版本,实际测量全部是 4.58 bit——同一种文件,挂 4 个不同的"低显存"招牌。

行业怎么看

审计者本人的语气克制——他强调这是 llama.cpp 自 2023 年以来的设计行为,量化工具确实会在日志里打印警告。问题在于警告写进了量化日志,而绝大多数用户下载的是别人做好的成品 GGUF,根本看不到那条警告。

但社区里也有反对声音。有人指出这不是"工具 bug",而是"信任链 bug"——模型卡(model card,模型说明页)里写的精度、文件名写的精度、metadata 里描述的精度,三方一致地指向了用户根本拿不到的那个版本。当 bartowski(业内最受信任的量化师之一)等头部玩家都中招时,问题已不是某一家粗心,而是整个本地 AI 生态缺乏文件级校验机制。

另一个值得警惕的风险是:这次审计只覆盖 25 个仓库,而 Hugging Face 上的 GGUF 数量是几万量级。我们有理由相信 14.4% 这个数字很可能被严重低估——这是抽样偏差,不是定论。

对普通人的影响

对企业 IT:如果公司正在评估"用开源模型本地部署降本",那么"这个 2-bit 模型只要 16G 显存就能跑"这类说辞,从今天起要先校验再信,否则算力预算可能算崩。

对个人职场:用 Mac 笔记本或消费级显卡跑大模型的从业者——程序员、咨询顾问、研究助理——如果你最近发现"明明说省显存,实际还是跑不起来",现在大概知道为什么了。

对消费市场:厂商宣传的"AI 笔记本本地跑 70 亿参数"在 2025 年密集出现,量化精度的透明度问题最终会传导到消费端——买回来发现跑不动,锅不在电脑,在上游文件注水。