A larger reasoning system inside an embedded edge envelope

Reinventy has promoted NVIDIA Nemotron 3.5 Lightning 30B-A3B NVFP4 as the canonical reasoning engine in Odino 3.11.0/Q19 on NVIDIA Jetson AGX Thor. TensorRT Edge-LLM 0.10.0 serves the production model; the former Nano engine is inactive and no automatic fallback or model selector is enabled.

Nemotron 3.5 is a 30-billion-parameter mixture-of-experts model with approximately 3 billion parameters active per token. It provides 7.5 times the total parameter capacity of the former Nano-4B engine while retaining a contained active-compute footprint suited to an accelerated embedded platform.

Production performance

Against the sustained comparable Nano baseline, Odino 3.11 records:

  • 72.623 tok/s across the five-run production HTTP mean, an improvement of 27.6%;
  • 77.104 tok/s on the qualified fixed HTTP workload, an improvement of 35.5%;
  • 92.8 tok/s at native-engine peak, an improvement of 63.1%;
  • 68.072 ms median time to first production content.

The production engine occupies 19,102 MiB of GPU unified memory at full residency. A supervised warm-engine restart returns readiness in four seconds.

Scientific reasoning and direct evidence

On the identical fixed bilingual scientific and adversarial manifest, measured accuracy increased from 63.33% to 100%. Direct Evidence p50 fell from 3,556.867 ms to 31.370 ms, and the former catalogue request that exceeded the HTTP limit now completes successfully with a grounded tool trace across 236 catalogue rows.

Knowledge Graph, claim-level provenance, evidence assessment, local uncertainty and out-of-domain handling, grounded tools and specialist read-only consumers remain integrated without enabling mutation, export, ingest or autonomous experiment execution.

The milestone demonstrates Reinventy’s core edge thesis: more capable reasoning does not need to imply cloud dependency, hidden authority or unbounded execution. Performance, evidence and governance can advance together on one NVIDIA accelerated edge node.

Sources: Odino 3.11.0 / Q19 governed production record · verified 2026-08-23