| Access today | ||||||
|---|---|---|---|---|---|---|
| Loading 3,651 models… | ||||||
Figures: plain = reported in published research data · HF from the model's Hugging Face weights · est. calculated from size and data or hardware · rough typical ratio for its field and year, within about 10× · – not disclosed by the developer.
Select a row to see its full details. Each model shows its verification level: L0 needs a data fix, L1 passes automatic checks, L2 is source-verified by a reviewer, L3 is fully verified including India hosting.
Only you, as the Atlas owner, can see this. Each row is a sales lead.
| Received | Contact | Model | Score | Serious findings | Notes |
|---|
Visible only to you.
| Date | Model | Score | Verdict |
|---|
Each model is placed by the strongest signal available. Steps run in order; a model stops at the first one that applies. Select a count to list those models.
| Step | Rule | Confidence | Models |
|---|
| Model | Organisation | Released | Current tier | Access today |
|---|
Red cells are models whose weights are already public although their proposed tier says they should be controlled or hosted only. Low-confidence precautionary tiers are counted in the grid but not flagged as released beyond their tier. Select any number to list those models.
| Proposed tier | Open weights | API or hosted | Unreleased | Not recorded | Total |
|---|
| Model | Organisation | Released | Proposed tier | Compute |
|---|
Thresholds: 10²⁵ FLOP is where the EU AI Act presumes systemic risk for general-purpose models. 10²⁶ FLOP and current frontier status mark the hosted-only tier proposed here. Parameters stand in for compute only when compute isn't reported.
| Developer | Models | Below L2 | Missing figures | Risk score |
|---|
Models with neither figure, by why they're missing.
| Group | Models | Action |
|---|
Fix these to reach Level 1.
| Check | Models |
|---|
Every sign-off, figure update and tier review, newest first. This is the audit trail.
| When (UTC) | Action | Model | Detail |
|---|
Open-weight models in the controlled or hosted tiers, largest first. These are where a break-it test (can the safety be fine-tuned away within a fixed compute budget?) matters most.
| # | Model | Organisation | Released | Compute | Tamper test |
|---|
Train models so removing safety by fine-tuning costs far more compute than it's worth.
Atlas field to add: compute an attacker needed to break it.
Remove specific dangerous knowledge, such as pathogen enhancement, from the weights so there is nothing to unlock.
Atlas field to add: unlearning applied and verified.
Keep the most dangerous material out of training from the start. The most robust option available now.
Atlas field to add: hazardous-data filtering disclosed.
Red teams with a fixed compute budget try to fine-tune safety away. Open release only if they fail.
Atlas field to add: test result, tester and date.
A public–private fund with frontier labs contributing a share of training spend, prize challenges for anyone who breaks a "tamper-resistant" model, and shared benchmarks so methods can be compared.
Screening DNA-synthesis orders, controlling access to lab equipment and chemicals, and using AI to patch software vulnerabilities faster than attackers exploit them.
Tiers manage the risk now. Once a model passes verified tamper-resistance tests, it can move down a tier and be released more openly.
| Year | Models | Open weights | Frontier |
|---|
| Organisation | Models | Open weights | Frontier | Latest model |
|---|
| Country | Models | Open weights | Frontier |
|---|
A model built by organisations in several countries counts once for each. About a third of models have no country recorded. Select a country to list its models.
The Atlas covers notable models: those that were state of the art, highly cited, widely used or otherwise significant, rather than every model published. Fields with no published figure show as "Not reported". Check a provider's current terms before relying on any entry.
Data sources: model records adapted from Epoch AI, Data on AI Models (CC BY 4.0), combined with Lekha's own checks, developer disclosures and public model files.