01 · COGNITIVE CLUSTER
Tin Man v2.1
12-core multimodal cluster on NVIDIA Blackwell
First publicly documented Nemotron-3-Nano-Omni V3 edge deployment running 100% NVFP4 native on NVIDIA Jetson AGX Thor 5000 — visual and audio encoders included. Prefill throughput remains constant from 8K to 100K tokens, a Hybrid Mamba-2 + Transformer GQA characteristic that pure transformer architectures cannot achieve at this scale on edge hardware.
SPRINT 2 CLOSED · EXTERNAL-DISTRIBUTION-READY REPORT
View specifications →