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...
Context
262Ktokens
About 374 pages of text at once
Max output
183Ktokens
About 261 pages in one answer
Intelligence
11.8/ 100
Above 16% of the catalogue
Output speed
69tokens/s
A 500-word answer in about 10.1 s
From public market statistics for this model
First token
1.3s
Reasoning models start later
From public market statistics for this model
Health, 30 days
99.97%
25 of 30 days without a fault
IranRouter’s measurements
Price & cost
Every rate in Rial. A request pays only for the tokens it actually uses.
| Input | 1,644,596 | IRR / 1M tokens |
|---|---|---|
| Output | 6,578,383 | IRR / 1M tokens |
| Reasoning tokensThe output rate | 6,578,383 | IRR / 1M tokens |
| Cache read | 328,919 | IRR / 1M tokens |
With prompt caching, the fixed part of your prompt is billed at the cache-read rate on later requests.
What would this model cost you a month?
Monthly cost
266.42MIRR
8.88M IRR a day, 4,440 IRR each request
The same work on peer models
30 days at today’s rate.
Rial price over 90 days
It moves with the exchange rate; the steps are the maker’s own price changes. Built at today’s margin, so approximate if the margin changed in the window.
Performance, 30 days
Health is measured by IranRouter itself: a day counts as healthy unless requests to this model failed on our side or it could not be served. First token and speed come from IranRouter’s traffic, and from public market statistics for this model on days without enough of it.
Daily health99.97%
Time to first token1.61 s
Output speed64 tokens/s
Place in the market
By usage across the global model market.
Daily rank#16
Share of the day’s tokens1.4%
Among its peers
Cost of one message against intelligence; this model is highlighted.
Cost of one message, IRR (log scale)
Capabilities
- Input
- Text
- Output
- Text
Parameters it accepts
frequency_penaltyinclude_reasoninglogit_biasmax_tokensmin_ppresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_pOther benchmarks
Design contests
- Code (overall)Elo: 1,145Win rate: 36%
- WebsitesElo: 1,145Win rate: 35%
- Data vizElo: 1,143Win rate: 38%
- Game devElo: 1,139Win rate: 36%
4 moreShow less
- UI componentsElo: 1,135Win rate: 36%
- 3DElo: 1,127Win rate: 38%
- ASCII artElo: 1,098Win rate: 37%
- SVGElo: 1,068Win rate: 32%
Standard tests
- GPQA DiamondScore: 81%Tasks: 172
- τ-bench Verified AirlineScore: 72%Tasks: 50
Start in a minute
The official OpenAI and Anthropic SDKs; only the address and the key change.
https://iranrouter.com/v1nvidia/nemotron-3-ultra-550b-a55b# pip install openai from openai import OpenAI client = OpenAI( base_url="https://iranrouter.com/v1", api_key="ir-...", ) r = client.chat.completions.create( model="nvidia/nemotron-3-ultra-550b-a55b", messages=[{"role": "user", "content": "Hello, introduce yourself"}], ) print(r.choices[0].message.content)
Peer models
Its family and models with a close score; the last column compares one message’s cost with this model.
| Model | Against this model | ||||
|---|---|---|---|---|---|
| Google: Gemini 3.5 Flash LiteGoogle | 5,069 | 11.7 | 132 | 1M | 23% cheaper |
| Meta: Muse Glimmer 30BMeta | 3,288 | 6.3 | 81 | 131K | 50% cheaper |
| MoonshotAI: Kimi K3Moonshot AI | 21,602 | 17.4 | 63 | 1M | 228% dearer |
| NVIDIA: Nemotron 3.5 LightningNVIDIA | 658 | 5.5 | 108 | 262K | 90% cheaper |
| MiniMax: MiniMax M3Minimax | 3,288 | 18.7 | 87 | 1M | 50% cheaper |
Technical specs
- ID
nvidia/nemotron-3-ultra-550b-a55b- Maker
- NVIDIA
- Tokenizer
- Other
- Released
- Jun 4, 2026
- Added to IranRouter
- Aug 7, 2026
- Context
- 262,144 tokens
- Max output
- 182,520 tokens
Default parameters
top_p=0.95temperature=1Scores by Artificial Analysis, Design Arena, as of Oct 31, 2026