Ctrl K
Back to models

NVIDIA

Nemotron 3 Ultra 550B A55B

  • Released
  • Open weights
  • OpenMDW
Compare
Intelligence Index

Reasoning setting

22.9#47 of 83measured models
Measured · Artificial Analysis v4.3.2
API price from
$0.50 in / $2.20 out / 1M tokOpenRouter · DeepInfra · FP4
Context window
262.14K tok

Performance profile

ReasoningRank #31 of 107
86.7%
GPQA Diamond
CodingRank #62 of 94
40.3%
SciCode
Cost efficiencyRank #44 of 67
39.7pts / $
Intelligence ÷ eval cost

Capabilities

Context window

Modalities

Measured setting

Integration

Reasoning trade-offs

Source (opens in a new tab)
Intelligence Index v4.3.2
Nemotron 3 Ultra 550B A55B
Cost per task (USD) · log scale
Nemotron 3 Ultra 550B A55BReasoning
Intelligence Index
22.9
Cost / task
$0.577
Evaluation time
3.1m

Pricing & access

API route
OpenRouter · DeepInfra · FP4API
Tier & contextStandard

USD per 1M tokens

Input
$0.50
Output
$2.20
Cached input
$0.10

Run locally

GGUF
Context
Estimated memory

~368.5 GiB

RAM + VRAM

Estimate assumptions

One text chat at the selected context, with a 16-bit cache and app allowance. Operating system and image processing are extra; optimized caches can use less.

384 GiB RAMwithout GPU offload

Memory budgetOne text chat
Weights
334.6 GiB
Context + state
0.5 GiB16-bit cache
App allowance
33.5 GiB
Download
359.2 GB9 files · keep together
Disk space
~401 GB
Available software
Ollama (opens in a new tab)LM Studio (opens in a new tab)Jan (opens in a new tab)KoboldCpp (opens in a new tab)llama.cpp (opens in a new tab)
Get GGUF (opens in a new tab)
Set up in Ollama · Q4_K_M · 8K context

Save the following as Modelfile. The context matches the estimate above.

FROM hf.co/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF:Q4_K_M
PARAMETER num_ctx 8192

Then run these commands in your terminal:

ollama create llmcurrent-local -f Modelfile
ollama run llmcurrent-local

Ollama import instructions (opens in a new tab) · For LM Studio or Jan, open the app’s model search, enter unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF, choose Q4_K_M, and set context to 8,192 tokens in the app’s model settings. Software links above open each app’s setup guide.

vLLM setup (opens in a new tab)

Benchmarks

Full leaderboardSource (opens in a new tab)

Humanity’s Last Exam

Frontier reasoning

Nemotron 3 Ultra 550B A55B · Reasoning#54 of 126

  1. 1Claude Opus 5.5Max61.4%
  2. 2Claude Fable 5.1Max59.1%
  3. 3Gemini 4 ArgonHigh57.1%
  5. 51Solar Pro 4Reasoning29.2%
  6. 52Qwen3.5 397B A17BReasoning29.0%
  7. 53Solar Open2 250BReasoning28.5%
  8. 54Nemotron 3 Ultra 550B A55BReasoning28.4%
  9. 55Qwen3.5 122B A10BReasoning25.2%
  10. 56Nemotron 3.5 Bolt 74B-A9BReasoning24.7%
  11. 57Gemma 4 31BReasoning23.6%

Peers use their best measured result on this benchmark.

Alternatives

Each model at its highest-scoring measured setting · Intelligence Index v4.3.2 and its evaluation cost per task.