Breaking

NVIDIA Releases 550B MoE Model Nemotron 3 Ultra for Long-Running Agents

June 9, 2026 at 14:00 EDT

Foundation Models
AI Agents
Open Source

In June 2026, NVIDIA released Nemotron 3 Ultra, an open-weight large language model designed for long-running AI agents. Built as a Mixture-of-Experts (MoE) model with 550B total parameters (55B active), its weights are distributed on Hugging Face, with integration into Perplexity and one-click deployment on AWS SageMaker.

June 2026 · NVIDIA

Nemotron 3 Ultra: A Fully Open Frontier Model Built for Long-Running Agents

NVIDIA ships a 550B Mixture-of-Experts model (55B active) with weights, training data and recipes all published — engineered for speed in agentic workflows, and positioned as a U.S. answer to the rise of Chinese open-weight models.

550B

total parameters
55B active (MoE)

token context length

~90%

MoE sparsity for efficient inference

300+

tokens/sec (BF16)
vs 50–100 typical

Throughput vs Rival Open Models

Relative generation speed at 8k input / 64k output. Each column = Nemotron 3 Ultra's speed multiple.

5.9×

vs GLM-5.1
754B

4.8×

vs Kimi-K2.6
1T

1.6×

vs Qwen-3.5
397B

What Powers the Speed

Hybrid Mamba-Transformer MoE

→

Multi-Token Prediction + speculative decoding

→

NVFP4 low-precision pretraining

Scored 48 on the Artificial Analysis Intelligence Index — among the top U.S.-origin open-weight models.

Strengths

Fast, cost-efficient agentic tool calls and coding
Fully open: weights, data and recipes published
Light censorship, high practicality
Day-one support: NIM, vLLM, Ollama, Unsloth

Caveats

Raw intelligence can lag models like GLM-5.1
Quality varies with presets and quantization
High throughput favors the NVIDIA stack
Guardrail behavior may shift over long contexts

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.