Breaking

NVIDIA Releases Open MoE Model Nemotron 3 Ultra: 550B Total, Up to 5x Faster Inference

June 16, 2026 at 14:01 EDT

Foundation Models
Open Source
AI Agents

On June 4, 2026, NVIDIA released Nemotron 3 Ultra, a Mixture-of-Experts model with 550B total and 55B active parameters, as an open model. Designed for long-running AI agents, it ships with full weights, training data and recipes openly available.

June 2026 · NVIDIA

Nemotron 3 Ultra: a 550B Open-Weight Engine Built for Long-Running Agents

NVIDIA's flagship Mixture-of-Experts model activates just 55B of 550B parameters, handles up to 1M tokens, and ships with open weights, open data, and open recipes — aimed at planning, tool-calling agents run in-house.

550B / 55B

Total / active params (MoE)

Max context (tokens)

−30%

Agentic task cost

Throughput vs. open frontier models

Relative speed at 8k-input / 64k-output (×)

5.9×

vs GLM-5.1
754B

4.8×

vs Kimi-K2.6

1.6×

vs Qwen-3.5

Attributed to the Blackwell architecture and NVIDIA's 4-bit NVFP4 format.

91%

PinchBench (agentic)

67%

Terminal-Bench 2.0

38–47.7

Intelligence Index

Favorable in practice

Code review quality on par with baseline (56–62% pass) at lower latency
Legal task reached Sonnet/Opus-level rubric (83%) at lower cost after post-training
Promising foundation for long-running agents

Practical caveats

Lower all-pass rates without post-training
Memory & parallelism settings need troubleshooting locally
Speed limits / lower priority on free tiers; data-usage terms to watch

The open question

How far can self-hosted agentic AI be put into practice?

Local runs peaked at 42–43 tokens/sec at 512K context on four DGX Spark (GB10/Blackwell) units with SGLang — built for enterprises seeking cost and control beyond API dependence.

Continue reading

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