The trust score
The trust score runs from 0 to 100. It is not a safety test. It measures how clear and open a model is. We add up six parts. Each part uses facts we can link to.
| Part | Points | How we give points |
|---|---|---|
| Official source | 25 | The files come from the maker's own account. |
| License | 25 | 25 for an OSI open source license. 12 for a custom license that allows use. 0 for no business use. |
| Business use | 15 | 15 with no caps. 8 with limits. 0 if not allowed. |
| Open data | 15 | 15 if the maker shares its training data and recipe. |
| No sign-up | 10 | 10 if anyone can download it. 5 if you must sign up first. |
| Facts checked | 10 | Share of our facts we could confirm from a source. |
A low score does not mean a model is bad or unsafe. It means it has more rules, or we could check less about it.
The quality tier
We do not run our own tests yet. So the quality tier is based on size. Bigger models usually give better answers. Newer models often beat older ones of the same size.
- Basic: under 4 billion weights
- Good: 4 to 10 billion
- Strong: 10 to 20 billion
- Very strong: 20 to 50 billion
- Top home tier: over 50 billion
The RAM estimate
A 4-bit file takes about 0.6 GB for each billion weights. To get the RAM you need, we take the file size, add a fifth, then add 2 GB for the app. It is a rough guide. Long chats use more memory.
How we check facts
We read the model page and license from the maker or its Hugging Face account. Each fact in our data has a source link, a date, and a status. A fact is "verified" if we read it from that source on that date. Anything else says "not yet checked". You can get all of it in our JSON file.
Common questions
What is a trust score?
A 0 to 100 score for how clear and open a model is. It adds points for an official source, an open license, business use, open data, no sign-up, and checked facts.
Does a high trust score mean a model is safe?
No. It means the license and source are clear. It is not a safety or quality test.