Provenance Record

Clear, well-scoped model-launch News piece (654 words) that separates AWS's vendor framing from MarkTechPost's architecture/benchmark detail and is careful to flag the throughput figures as NVIDIA-measured.

Both snapshots read from disk (HTTP 200). source-0 (AWS SageMaker JumpStart blog) confirms — from its spec table and lead — the 'Hybrid Transformer-Mamba MoE' architecture, '550B total / 55B active', 'Up to 1M tokens' context, 'Text in, text out', NVFP4 precision, and the '5x faster inference and up to 30% lower cost for agentic workloads' claims, plus day-zero SageMaker availability. source-1 (MarkTechPost) confirms the 550B/55B totals, 108 layers, 8,192 model dimension, 512 experts with top-22 routed, the NVFP4 E2M1 4-bit datatype with two-dimensional block quantization, the 94.7 RULER score at 1M tokens, the '8K input / 64K output … on GB200, … 5.9x the throughput of GLM-5.1' and '4.8x faster than Kimi-K2.6' figures, the Terminal Bench split (Nemotron 56.4 vs Kimi-K2.6 leading at 67.2), the OpenMDW-1.1 license, and the Hugging Face / NVIDIA NIM / Nebius / OpenRouter / Together AI / Perplexity availability list.

Every fabrication-risk specific (550B/55B, 108/8,192/512/22, E2M1, 94.7 RULER, 5.9x-vs-GLM-5.1, 4.8x-vs-Kimi, 67.2/56.4 Terminal Bench, OpenMDW-1.1) is verbatim in a cited source with the correct numeric pairing. The 5.9x↔GLM-5.1 and 4.8x↔Kimi-K2.6 pairings match the source exactly, and the benchmark direction (Kimi leads accuracy, Nemotron leads throughput) is reported correctly. Two-source piece, but both are reputable and the headline 55B/550B figures and architecture appear in both.

Fully-verified open-weights model launch with disciplined numeric attribution and correct benchmark framing. APPROVE.