The production brain behind the Tin Man OS product line
Reinventy has completed the Tin Man V5.47.0 production cutover for Nemo, the multimodal front door of its 26-core sovereign cognitive fleet on NVIDIA Jetson AGX Thor.
The promoted TensorRT Edge-LLM 0.9.1 runtime reaches 98.22 tok/s decode in the matched batch-1 qualification, compared with 39.01 tok/s on the prior runtime—a 2.52× uplift. Prefill increases from 1,638.0 to 9,945.4 tok/s, a 6.07× uplift, while the LLM engine memory footprint falls by 1,262 MiB.
A separate 511-token probe sustains 93.74 tok/s with 93–97% GPU utilization. Text, SSE streaming, vision, short and long audio, and supervised crash recovery all passed the production gate.
Tin Man is not one model endpoint. Its 26 live cognitive cores coordinate perception, dialogue, memory, science, health, operations and governed execution inside one visible human-authority boundary. Camera and LIDAR perception, streaming voice, multilingual speech and specialist reasoning share the same embedded compute substrate.
This production lineage matters directly to Tin Man OS. The operating-system product line does not begin from a speculative architecture; it productizes the hard-won integration patterns of a sovereign synthetic brain already operating on NVIDIA accelerated edge hardware. The published Tin Man OS alpha does not yet claim the runtime performance of V5.47, but V5.47 provides the engineering lineage the product line is designed to carry into reproducible, hardware-aware editions.
Sources: Tin Man V5.47.0 governed production record · matched qualification verified 2026-08-22