Nemotron 3 UltraOpen Source

Released

Nemotron 3 Ultra is an AI model released by NVIDIA on Thursday, Jun 4 2026, 85 days after Nemotron 3 Super. It is an open-source model — weights, and training code/data, released under an open license. It is a 550B parameter model. Benchmark results (shown below) cover BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, BrowseComp, Humanity's Last Exam, GPQA Diamond, and 1 more.

Available from

DeepInfra$0.50$2.20256Kfp4
BaseTen$0.60$2.40200Kfp4
Venice$0.625$3.125256Kfp8
USD per 1M tokens

Benchmarks

Coding

SWE-Bench Verifiedvia BenchLM
71.9%
#36 of 57

Terminal & CLI

Terminal-Bench 2.1via BenchLM
56.4%
#30 of 32

Agentic & tool use

BrowseCompvia BenchLM
44.4%
#28 of 28

Reasoning & science

Humanity's Last Examvia BenchLM
26.7%with tools
#43 of 46
GPQA Diamond
86.7%
#26 of 59

Knowledge work

AA Intelligence Index
48
#5 of 6

Robustness

BullshitBench v2
49%
#37 of 94

Source: BenchLM, retrieved 28 September 2026. Other sources are identified on the linked benchmark pages.

Compare Nemotron 3 Ultra with

Nemotron 3 Ultra

About

Nemotron 3 Ultra, announced at Computex on June 1, 2026 and released with weights on June 4, was NVIDIA's largest published model: 550 billion total parameters with 55 billion active, in the same hybrid Mamba-Transformer mixture-of-experts architecture as its smaller siblings. It completed the Nano/Super/Ultra ladder the company had laid out when the generation opened the previous December.

At release it scored 48 on the Artificial Analysis Intelligence Index — the highest of any US-developed open-weight model at the time, well clear of the previous US leader Gemma 4 31B at 39, though still behind the strongest Chinese open flagships on the same index — alongside 86.7% on GPQA Diamond. NVIDIA leaned on the efficiency argument it had built the family around: over 300 tokens per second on its own hardware and markedly lower per-task cost than the open models above it on the leaderboard.

Frequently asked questions

Nemotron 3 Ultra was released by NVIDIA on Thursday, Jun 4 2026.

All NVIDIA releases

9 tracked