Supported models
Which model families can be scanned, what each can detect, how the base model is found, and stored results.
Integrity Lens scans Hugging Face models in safetensors format, up to 20 GB, that belong to a supported family: fine-tunes, merges and edits of the base models below. Public models need nothing; private and gated models need a Hugging Face token.
Families
A model belongs to a family when its architecture and size match the base (config.json). Each family has a level:
| Level | What it measures | Verdicts |
|---|---|---|
| C calibrated | weights and activations, with a calibration curve | Intact / Degraded / Removed, with confidence and predicted harmful-request rate |
| B activations | weights and activations | Abliterated / Review / Not detected |
| A weights | weights only | Abliterated / Review / Not detected |
Give a verdict today (validated against known strips and benign fine-tunes):
| Family | Base | Level |
|---|---|---|
| Gemma 3 4B | google/gemma-3-4b-it | C |
| Gemma 3 1B | google/gemma-3-1b-it | C (activations decide; the weight check is not validated) |
| Llama 3.1 8B | meta-llama/Llama-3.1-8B-Instruct | B |
| Mistral 7B v0.3 | mistralai/Mistral-7B-Instruct-v0.3 | B |
| Qwen3 4B | Qwen/Qwen3-4B | B |
| Llama 3.2 3B | meta-llama/Llama-3.2-3B-Instruct | A |
| Qwen2.5 0.5B | Qwen/Qwen2.5-0.5B-Instruct | A |
| DeepSeek R1-Distill-Qwen 7B | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | A |
Measured, verdict pending validation (results are not_assessed, with the measurements as evidence): Gemma 2 2B
and 9B, Qwen3 0.6B and 1.7B, DeepSeek R1-Distill-Qwen 1.5B, MiniCPM 1B, HY-MT1.5 1.8B, LFM2.5 1.2B.
Above the 20 GB checkpoint limit (not scannable yet): Gemma 3 12B (validated), Mistral Nemo 12B, Mistral Small 22B, Phi-4, Qwen3.5 27B.
The list grows as families are calibrated. The current one, with each family's validation notes:
curl -s https://api.neuronlens.com/v1/integrity/families -H "X-API-Key: $NEURONLENS_API_KEY"
A model from an unsupported family returns not_assessed at once and is not counted.
How the base model is found
From the model card's base_model, then config.json, then the model's name. The result shows what was used
(measured_against). If the card names the wrong base, pass the right one as compare in
POST /v1/integrity/scans, or as parent in the one-call
POST /v1/integrity/scan.
Stored results
Results are stored by model and revision (the commit on Hugging Face). Scanning a revision that was already
measured returns at once. A new commit is a new revision and is measured again. Pin the revision you ship
(hf://org/name@<commit>) so your result always refers to exactly those weights.