Models
A repository is not a deployment
Model hubs list repositories. Devices load files. LokiAI closes that gap and records how it did so.
- Hugging Face integration
- GGUF
- TFLite
- Compatibility checks
01Verification
Every deployment names one file
Discovery may return dozens of quantisations. Only one is deployed, and the record says which.
| Check | Resolved value | Verdict |
|---|---|---|
| Repository | hf.co/<org>/<model> | Resolved |
| Exact file | model-q4_k_m.gguf | Pinned |
| Format | GGUF v3 | Supported |
| Size | 1.94 GB | Fits headroom |
| Runtime | llama.cpp adapter | Available |
| Hardware fit | arm64 · 8 GB RAM | Compatible |
| Download URL | CDN direct | Reachable |
02Formats
What LokiAI can load today
GGUF
Quantised language models via the llama.cpp adapter. Available on Android and the beta Linux agents.
TFLite
Detection, classification and audio models. CPU everywhere, NNAPI where the device exposes it.
More formats
Under evaluation. Not deployable yet — listed so the direction is clear.
03Compatibility
Checks that run before a single byte is downloaded
- Artifact size against reported free storage and memory headroom.
- Format against the runtimes the agent has actually registered.
- Architecture and OS version against the runtime's requirements.
- Accelerator availability, when the runtime can use one.
- Source reachability, so a broken URL fails fast instead of mid-transfer.