Nemotron 3 Ultra
Nemotron 3 Ultra is NVIDIA's largest open reasoning model, a hybrid Mamba-Transformer MoE with 550B total and 55B active parameters, latent MoE routing, multi-token prediction, and a context window of 1M tokens for long-running agent workflows.
import { streamText } from 'ai'
const result = streamText({ model: 'nvidia/nemotron-3-ultra-550b-a55b', prompt: 'Why is the sky blue?'})About Nemotron 3 Ultra
NVIDIA released Nemotron 3 Ultra on June 4, 2026 as the largest model in the Nemotron 3 family, completing the tier above Nano and Super. It carries 550B total parameters with 55B active per token, and NVIDIA positions it as the reasoning and orchestration layer for long-running agent workflows: the model that handles planning, synthesis, and verification while lighter models execute routine steps.
The architecture interleaves three layer types. Mamba layers process long sequences with linear-time complexity, which keeps a context window of 1M tokens practical. Transformer attention layers appear at select depths to preserve precise recall from large contexts. Latent mixture-of-experts (MoE) routing compresses token embeddings into a smaller latent space before selecting experts, so distinct specialists activate for reasoning, coding, and tool calls without dense compute. Multi-token prediction (MTP) layers predict several future tokens per forward pass, providing built-in speculative decoding for long outputs.
Nemotron 3 Ultra scores 91% on PinchBench, 82% on IFBench, and 95% on Ruler at 1M tokens. Weights, data, and recipes are released under the Linux Foundation's permissive OpenMDW-1.1 license. Full details: https://www.together.ai/models/nvidia-nemotron-3-ultra.