Breaking

NVIDIA Releases Nemotron 3 Models, Open Weights Up to 550B

June 23, 2026 at 14:00 EDT

Foundation Models
Open Source
AI Agents

NVIDIA has released Nemotron 3, a family of open-weight large language models built on a hybrid Mamba-Transformer MoE architecture, rolling out three lines: Nano, Super and Ultra. The largest model, Ultra (550B parameters), had its final release on June 4, 2026, with full-family details laid out in a "Nemotron Office Hours" technical walkthrough and Q&A.

June 4, 2026 · NVIDIA

Nemotron 3: Open-Weight LLMs Built for Speed on NVIDIA GPUs

NVIDIA's new open family — Nano, Super and Ultra — pairs a hybrid Mamba-Transformer MoE core with a 1M-token context, publishing weights, data and recipes to chase efficiency against China's open giants.

31.6B

Nano total params · 3.2B active · single-H200 local-friendly

~110B

Super total params · ~10-12B active · mid-scale balanced

550B

Ultra total params · 55B active · agentic / long-context

Token context window · Mamba-2 + Transformer attention

Throughput edge — Ultra vs rivals

At 8K input / 64K output, Ultra runs multiples faster

1×

baseline

1.6×

vs Qwen-3.5

4.8×

vs Kimi-K2.6

5.9×

vs GLM-5.1

Nano shows a similar story on a single H200: 3.3× Qwen3-30B-A3B and 2.2× GPT-OSS-20B at 8K / 16K.

What makes it efficient

Hybrid Mamba-Transformer MoE → LatentMoE + MTP layer → NVFP4 training → Reasoning budget control

High-transparency release: open weights, training data and recipes, commercial use permitted. Runs on NIM, vLLM, SGLang, Ollama, llama.cpp and Amazon SageMaker JumpStart.

Developer praise

Nano: outstanding compute efficiency, strong at coding, debugging and summarization, fast locally
Ultra: completes long agentic tasks quickly, lowers inference cost

The caveats

Super seen as modest relative to expectations
Ultra is resource-heavy for local inference — large GPU clusters recommended

Continue reading

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