Models¶
Moveris offers five model variants (Mixed V2) to match different integration needs. All models use the same architecture, trained on 6,486 videos across 9 attack types — physical spoofing, deepfakes, screen replay, faceswap, and others. The difference is how many frames are analyzed, which affects latency and temporal coverage.
| Fast | Balanced Recommended | Thorough | Extended | Maximum | |
|---|---|---|---|---|---|
| Model ID | mixed-10-v2 | mixed-30-v2 | mixed-60-v2 | mixed-90-v2 | mixed-120-v2 |
| Frames | 10 | 30 | 60 | 90 | 120 |
| Capture time | ~0.3 s | ~1 s | ~2 s | ~3 s | ~4 s |
| Response time | ~1 s | ~3 s | ~5 s | ~7 s | ~10 s |
| EER | 4.4% | 4.0% | 4.7% | 5.7% | 4.6% |
| AUC | 0.988 | 0.991 | 0.991 | 0.989 | 0.991 |
| Balanced accuracy | 95.2% | 95.7% | 95.2% | 94.3% | 94.7% |
| Best for | Low-friction, high-volume flows | Standard KYC & identity verification | High-security onboarding | Escalation, compliance-heavy flows | Highest scrutiny, regulatory edge cases |
Model cards
Detailed model cards with full metrics and evaluation data are available in the Developer Portal. Sign in to access them.
Not sure? Start with Balanced (mixed-30-v2)
It delivers the best overall accuracy (4.0% EER, 0.991 AUC) with only 1 second of capture — the right default for most integrations.
When to Use Each¶
Fast — Users are impatient, risk is low, or you're re-verifying someone already trusted.
Balanced — Standard identity verification, account creation, financial services.
Thorough — A single false accept would be costly. Users will tolerate a longer check.
Extended — Compliance-heavy flows, escalation when Thorough is borderline, or when regulators expect longer temporal coverage (~3 seconds).
Maximum — Highest scrutiny scenarios: regulatory edge cases, high-value transactions, or final-level escalation when you need maximum temporal redundancy (~4 seconds).
Escalation Pattern¶
You don't have to pick just one. Use Fast by default and escalate when needed:
User submits → Fast (mixed-10-v2)
→ High confidence → Done
→ Borderline → Balanced (mixed-30-v2) retry
→ Still borderline → Thorough (mixed-60-v2) final check
→ Escalation → Extended (mixed-90-v2) or Maximum (mixed-120-v2)
How Frames Affect Results¶
More frames give the model more temporal signal to analyze — more micro-expressions, more biological coherence data across time. All v2 models achieve strong accuracy (94–96% balanced), so the primary tradeoff is latency vs. temporal redundancy rather than a large accuracy gap.
All models use the same API endpoints. To select a model, pass the model ID in your request:
Frames minimum (recommended exactly)
The API requires at least the number of frames the model expects. mixed-10-v2 expects 10 frames, mixed-30-v2 expects 30, mixed-60-v2 expects 60, and so on. For predictable latency, send exactly the required number.
Deprecated Models¶
The legacy mixed models (mixed-10, mixed-30, mixed-60, mixed-90, mixed-120, mixed-150, mixed-250) are deprecated. Migrate to the corresponding mixed-N-v2 variants (e.g. mixed-10 → mixed-10-v2).
See Model Versioning & Frames Parameters for alias behavior, version pinning, and the exact frame rules per endpoint.