A purpose-built lightweight reasoning model for structured scenes. Runs on RK3588, Jetson, and Pi-class boards. Quantised, memory-aware, designed to read the Scene Contract — not free-form text.
Native input format. No prompt engineering, no JSON wrapping — the model is trained on structured perception output.
Compact natural-language descriptions of what's happening, suitable for logs, dashboards, and human handoff.
Maps scene + memory + policy into candidate intents and risk levels — feeds the validation envelope.
20–80 ms typical on RK3588. Quantised int8/int4 weights. Memory-aware streaming context.
Persists temporal summaries instead of raw history — sites accumulate context without unbounded growth.
Train large, ship small. Continuous distillation from supervisor models keeps the edge model current.
| Target | Variant | Memory | Latency |
|---|---|---|---|
| RK3588 | scenelm-edge · int4 | ~1.2 GB | 30–70 ms |
| Jetson Orin Nano | scenelm-edge · int8 | ~2.4 GB | 20–50 ms |
| Pi 5 + Hailo | scenelm-lite · int8 | ~0.9 GB | 40–100 ms |
| x86 + GPU | scenelm-full · fp16 | ~6 GB | 10–25 ms |
One model lineage, multiple deployment targets and specialisations.
Sub-1 GB, Pi-class. Basic intent + summarisation.
The default. RK3588 / Jetson Orin class.
Tuned for inspection, defects, and process loops.
Multi-modal — vision + LiDAR + radar + audio cues.
Long-horizon temporal reasoning. Site-wise priors.
Forward modelling. Predicts the next-state for proactive intervention.
SceneLM ships with VAOS. No cloud round-trip required.