Model Drop

Nemotron 3 Ultra

June 4, 2026 2 min read

Nemotron 3 Ultra, which NVIDIA announced at Computex on June 1, 2026 and shipped as open weights on June 4, 2026, is a 550B-parameter hybrid Mamba-Transformer mixture-of-experts model built to run the hardest calls in long-running agents.

Model: Nemotron 3 Ultra (nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Maker: NVIDIA
Ship date: June 4, 2026
Pricing: Free (open weights)
Context: 1M tokens
Where you can use it:
- Hugging Face weights, BF16 and NVFP4
- build.nvidia.com, NIM, Perplexity, OpenRouter
When to reach for it:
- Orchestrating long-running agents
- Long documents and codebases, up to 1M
- Skip it for top coding-agent accuracy
How it compares:
- Much faster than Kimi K2.6 and GLM-5.1
- Trails both on coding agents and web search
What makes it different:
- Weights, data and recipes all released
- Mamba-Transformer hybrid trained in NVFP4
Early reactions: Early users said it lands near Kimi K2.6 but runs much faster, while The Batch called it far quicker than rivals but not top-tier.
Read more: NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents, Hugging Face model card, Nemotron 3 Ultra technical report

The benchmarks

Scores come from NVIDIA's model card and technical report against GLM-5.1, Kimi K2.6 and Qwen3.5 397B (NVIDIA also printed MiniMax M2.7 and DeepSeek V4 Pro and Flash); NVIDIA doesn't label a reasoning mode, ran every model at its own card's recommended settings, and used internal scaffolding for BrowseComp and PinchBench.

NVIDIA's numbers

Terminal-Bench 2.1 (terminal agent tasks): 56.4%. −2.9 under GLM-5.1 (59.3%), −10.8 under Kimi K2.6 (67.2%), +6.5 over Qwen3.5 397B (49.9%).
PinchBench (OpenClaw coding agents): 90.0%. +8.8 over GLM-5.1 (81.2%), −0.2 under Kimi K2.6 (90.2%), +3.4 over Qwen3.5 397B (86.6%).
BrowseComp (hard web research): 44.4%. −15.0 under GLM-5.1 (59.4%), −16.9 under Kimi K2.6 (61.3%), +3.9 over Qwen3.5 397B (40.5%).
LiveCodeBench (competitive coding): 89.0%. +3.3 over GLM-5.1 (85.7%), −1.2 under Kimi K2.6 (90.2%), +9.7 over Qwen3.5 397B (79.3%).
GPQA Diamond (PhD-level science, no tools): 87.0%. +0.9 over GLM-5.1 (86.1%), −4.0 under Kimi K2.6 (91.0%), −0.1 under Qwen3.5 397B (87.1%).
Humanity's Last Exam (no tools) (expert-level questions): 26.7%. −0.5 under GLM-5.1 (27.2%), −8.1 under Kimi K2.6 (34.8%), −1.8 under Qwen3.5 397B (28.5%).
IFBench (instruction following): 81.7%. +5.1 over GLM-5.1 (76.6%), +8.0 over Kimi K2.6 (73.7%), +3.5 over Qwen3.5 397B (78.2%).
RULER (1M) (long-context recall): 94.7%. +4.6 over Qwen3.5 397B (90.1%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 22.9. −3.2 under GLM-5.1 (26.1), −4.1 under Kimi K2.6 (27.0), +4.5 over Qwen3.5 397B (18.4), +0.1 over MiniMax M2.7 (22.8).