NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model from NVIDIA, with 55B active parameters out of 550B total (MoE). Built on a hybrid Transformer-Mamba mixture-of-experts architecture, it...
At a glance
- Context window
- 262Ktokens
- Max output
- 183Ktokens
- Knowledge to
- —
- Released
- 2026
Capabilities
Benchmarks
- Intelligence index
- 23
- Coding index
- 49
- Agentic index
- 22
Design Arena
| Category | Elo | Win rate |
|---|---|---|
| models · 3d | 1,163 | 40.5 |
| models · dataviz | 1,156 | 38.4 |
| models · codecategories | 1,155 | 36.3 |
| models · gamedev | 1,155 | 36.9 |
| models · uicomponent | 1,151 | 37.2 |
| models · website | 1,145 | 34.5 |
| models · svg | 1,099 | 35.5 |
| models · asciiart | 1,097 | 36.5 |
Standard evaluations
| Benchmark | Score | Tasks |
|---|---|---|
| gpqa_diamond | 0.82 | 579 |
| tau_bench_verified_airline | 0.77 | 150 |
Source: Artificial Analysis, Design Arena · as of Sep 16, 2026
Performance
Uptime on IranRouter
4.9%
Reference uptime
100%
Median latency
5,232 ms
Throughput
75.1 tok/s
IranRouter uptime comes from our own automated probes; the reference figures are daily measurements of this model's upstream routes. Each card shows a 30-day average, and a day without data stays empty.
Popularity
- Current rank
- #9
- Rank change, 7 days
- 0
Daily tokens · 90 days
386.1B
Ranked by daily token volume across the global model market; the figure above the chart is the daily average.
Supported parameters
Parameters this model accepts; anything else is ignored.
Quick start
from openai import OpenAI
client = OpenAI(
api_key="ir-...",
base_url="https://iranrouter.com/v1",
)
resp = client.chat.completions.create(
model="nvidia/nemotron-3-ultra-550b-a55b",
messages=[{"role": "user", "content": "سلام!"}],
)
print(resp.choices[0].message.content)Only base_url and the key change; the rest of your code stays as it is.