safetyAI model classification
The 4C classification of AI models, live
Every model on the Hugging Face registry and the major API models, plotted by risk class on a live classification sky — Certification, Classification, Containment, Control.
Browser · macOS · Windows · Linux · ChromeOS
What it does
Safety gives each model a capability score from its model-card metadata and places it in one of ten risk classes, from I (inert) to X (uncontained). Classes VI and VII need evidence from the model card or a sourced register; VIII to X stay empty until evidence exists. Safety-stripped derivatives — abliterated, uncensored — sit below the certification plane whatever their class.
How it works
- Open safety.vims.com
- Pick a class or search a model (press /)
- Inspect a star for its evidence, lineage and class
FAQ
Is a high class a judgement that a model is dangerous?
No. A class records measured capability and evidence; an empty reserved class is not a prediction.
Where does the evidence come from?
Model-card metadata, the stated training compute, agentic benchmarks, and a register of sourced claims shown verbatim.
Can I use VIMS Safety in a browser and in VIMS Desktop?
Yes. Open it at its subdomain, or use the same surface inside the VIMS Desktop client — same key, same data.
How does VIMS Safety connect to the rest of VIMS?
It shares your identity from keys.vims.com and is reachable by agents through MCP and Flow, so it works alongside every other VIMS app.
Related
The 4C classification of AI models, live
Safety gives each model a capability score from its model-card metadata and places it in one of ten risk classes, from I (inert) to X (uncontained). Classes VI and VII need evidence from the model card or a sourced register; VIII to X stay empty until evidence exists. Safety-stripped derivatives — abliterated, uncensored — sit below the certification plane whatever their class.