550B total, 55B active per token (MoE) - NVIDIA’s from-scratch hybrid Mamba-2 + Transformer with LatentMoE routing and 5-token Multi-Token Prediction (MTP) speculative decode. 262K default context (up to 1M with env flags); RULER@1M 94.7.
Open weights under the Linux Foundation’s OpenMDW-1.1 license, downloadable on HuggingFace (BF16 + NVFP4 checkpoints).
Server-class only. NVIDIA’s stated minimum is 8x H200, 16x H100, or 8x B200/GB200. BF16 weights are ~1.1TB and the NVFP4 checkpoint ~275GB, so it does not fit a single RTX 5090, RTX 6000 Ada, or Mac Studio M3/M4 Ultra. Not workstation-fittable, so cloud-only for nearly everyone despite the open weights.
Benchmarks are NVIDIA self-reported: MMLU-Pro 86.8, GPQA 87.0, SWE-bench Verified 70.7. Independent eval (Artificial Analysis, Intelligence Index 47.7, #9 of 89) calls it the leading US open-weight model but ~6 points behind Kimi K2.6 on hard reasoning and coding. Treat the vendor throughput figures as claims pending community replication.