Nemotron 3 Ultra (550B-A55B)

Positioning

Nemotron 3 Ultra (550B-A55B) is NVIDIA's frontier-scale open-weight reasoning model (June 2026) — and notably the most capable open model from a US lab, released fully open (weights + recipes) under the Linux Foundation's OpenMDW-1.1 commercial license.

What stands out

The architecture is the interesting part: a "LatentMoE" hybrid interleaving Mamba-2, MoE, and select attention layers with multi-token prediction, trained on an NVFP4 recipe. The NVFP4 checkpoints make a 550B model unusually tractable on Blackwell hardware, and 1M context is supported (NVFP4-on-Blackwell; 262K in BF16). A fully-open US-lab model with permissive licensing is rare and valuable for teams that want provenance.

Honest caveats

All benchmarks are NVIDIA-reported (via their Nemo Evaluator SDK) and not independently verified. Per Artificial Analysis it still trails the leading Chinese open models on the intelligence index. At 550B it is datacenter-class; the smaller Nemotron 3 Super (120B) and Nemotron 3 Nano (30B) tiers are the locally-practical choices for most readers.

Verdict

Run it if you need a fully-open, commercially-licensed frontier reasoner with documented training recipes and you have Blackwell / NVFP4 infrastructure. Most local users should start with Nemotron 3 Super or Nano instead — same family, same openness, hardware you can actually buy.