June 5, 2026
NVIDIA Nemotron 3 Ultra: I Gave This 550B Open Model 10 REAL Tasks
NVIDIA dropped Nemotron 3 Ultra — a 550B-parameter (55B active) open Mamba-Transformer MoE built for long-running agents, with a 1M-token context window. Instead of reading the benchmark charts, I gave it 10 real tasks live on NVIDIA's free build.nvidia.com API — and wherever it wrote code, I executed it; wherever it did math, I verified by hand.
Architecture
- Hybrid Mamba-Transformer MoE — Mamba keeps long sequences cheap, expert routing gives frontier quality without frontier compute per token
- Trained with multi-teacher on-policy distillation (10+ specialist teachers) + 212B fresh tokens (GitHub code through late 2025)
- 1M-token context (95% on RULER)
The result
10/10 — from "explain MoE in a sentence" to an O(1) LRU cache with passing tests, strict JSON, tool calling, and hard math (verified = 686). Plus the honest rough edges: free-tier queue waits and token-hungry thinking mode.
Try it free → https://build.nvidia.com · Model: nvidia/nemotron-3-ultra-550b-a55b
Full 10-task run in the video above. Subscribe for more real open-model tests.