Nemotron 3 Ultra 550B-A55B

Released 2026-06 · reasoning · 1.0M tokens · 6 benchmarks

Editorial notes

Release oficial 4 junio 2026 (keynote Computex GTC Taipei 1 jun); weights en HF desde ~25 mayo. Flagship open-weight mas potente de Nvidia. Latent MoE hibrido (Mamba-2 + MoE + Attention selectiva), Multi-Token Prediction, pre-train NVFP4 ~20T tokens, 90% sparsity. 550B totales / 55B activos. License OpenMDW-1.1. >300 tok/s, min HW 8x B200. Scores del HF model card oficial (GPQA/HLE son no-tools; HLE with-tools 37.4; LiveCodeBench v6). NVIDIA no reporta AIME — math via IMOAnswerBench 88.6, IOI-2025 570, Apex 74.9. AA Intelligence Index ~48 (third-party).

Spec sheet

Company: Nvidia
Country: US
Type: reasoning
Release: 2026-06
Context: 1.0M tokens
Slug: nemotron-3-ultra

Benchmarks (6)

Reasoning 3

Coding 3

Cite this model

BibTeX · APA

BibTeX

@misc{frontier-nemotron-3-ultra,
  title  = {Nemotron 3 Ultra 550B-A55B},
  author = {{Nvidia}},
  year   = {2026},
  note   = {Frontier Benchmarks AI atlas. Accessed 2026-06-17},
  url    = {https://frontierbenchmarks.com/models/nemotron-3-ultra}
}

APA

Nvidia (2026). Nemotron 3 Ultra 550B-A55B [Large language model]. Frontier Benchmarks AI. Retrieved 2026-06-17, from https://frontierbenchmarks.com/models/nemotron-3-ultra

Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.

Battle vs another model ← All models More from Nvidia Methodology