# one cognition engine

SceneLM™ — edge-distilled reasoning

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.

SceneLM is one engine inside the Reasoning Runtime — not the runtime itself. It runs alongside rules, local LLMs, and hybrid policies. The orchestrator routes work to it when contextual reasoning over structured scenes is needed at edge latency budgets.
# what scenelm does

Structured-scene reasoning at edge speed

🗺️

Reads the Scene Contract

Native input format. No prompt engineering, no JSON wrapping — the model is trained on structured perception output.

📝

Scene summarisation

Compact natural-language descriptions of what's happening, suitable for logs, dashboards, and human handoff.

🎯

Intent & risk

Maps scene + memory + policy into candidate intents and risk levels — feeds the validation envelope.

Edge latency

20–80 ms typical on RK3588. Quantised int8/int4 weights. Memory-aware streaming context.

💾

Memory compression

Persists temporal summaries instead of raw history — sites accumulate context without unbounded growth.

🔁

Distillation pipeline

Train large, ship small. Continuous distillation from supervisor models keeps the edge model current.

# deployment targets

Where SceneLM runs

TargetVariantMemoryLatency
RK3588scenelm-edge · int4~1.2 GB30–70 ms
Jetson Orin Nanoscenelm-edge · int8~2.4 GB20–50 ms
Pi 5 + Hailoscenelm-lite · int8~0.9 GB40–100 ms
x86 + GPUscenelm-full · fp16~6 GB10–25 ms
# future variants

The SceneLM family

One model lineage, multiple deployment targets and specialisations.

🪶

SceneLM-lite

Sub-1 GB, Pi-class. Basic intent + summarisation.

SceneLM-edge

The default. RK3588 / Jetson Orin class.

🏭

SceneLM-industrial

Tuned for inspection, defects, and process loops.

🔀

SceneLM-fusion

Multi-modal — vision + LiDAR + radar + audio cues.

💾

SceneLM-memory

Long-horizon temporal reasoning. Site-wise priors.

🎯

SceneLM-predictive

Forward modelling. Predicts the next-state for proactive intervention.

Cognition you can quantise.

SceneLM ships with VAOS. No cloud round-trip required.

See the Reasoning Runtime Download weights