[whatmodel.sh](../../README.md) / Nemotron 3 Ultra 550B-A55B
# Nemotron 3 Ultra 550B-A55B
NVIDIA · released 2026-06 · 550B params (55B active) · openmdw
**score 68.3** ±0.7 · #95 of 732 · from 32 benchmark results across 5 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Nemotron 3 Ultra 550B-A55B among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃▆█▄▅●▄▄▂▁ #87 of 313 22% 7
coding ▁▁▁▃▅▇█▇●▄▃▁ #90 of 469 20% 6
chat & writing ▁▁▁▃▆▄▆█▆●▇▃ #84 of 402* 4.7% 1
math ▁▁▃▅█▆▄●▄▂▁▁ #113 of 431 6.2% 3
web search ▁▁●▅▄█▃▄·▁·▁ #75 of 84* 4.4% 2
agent tasks · #87 of 313 · 22% of its score
Beats [Grok 4.20 (0309 reasoning)](grok-4-20.md), [GPT-5.1](gpt-5-1.md), [o3-pro](o3-pro.md) here, though they rank higher overall. Trails [MiniMax-M2.7](minimax-m2-7.md), [Step 3.5 Flash (196B-A11B)](step-3-5-flash.md), [Nemotron 3.5 Bolt 74B A9B](nemotron-3-5-bolt-74b-a9b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Verified (maker) 70.7% ●····██·████ beats 0/6 hard · avg 21% 6.2% sharp · at its level
SWE-bench Multilingual (maker) 67.7% ●··▄▄···▄▄·█ beats 0/6 hard · avg 30% 4.4% at its level
Terminal-Bench 2.1 (maker) 56.4% █··▄●·▄····▄ beats 4/6 very hard · avg 13% 4.1% sharp · at its level
Humanity's Last Exam with tools (… 37.4% ●····▄·█·▄·▄ beats 0/5 hard · avg 15% 2.4% at its level
BrowseComp (maker) 44.4% ▄·●▄···▄··█▄ beats 1/6 hard · avg 18% 2.0% at its level
τ³-bench (maker) 70.9% ▄····▄·●·▄·█ beats 2/6 medium · avg 42% 1.9% at its level
τ³-bench telecom (tau2-bench ≥ v1… 92.9% ▄·●···▄·▄·█▄ beats 1/6 easy · avg 79% 1.2% maxed out
coding · #90 of 469 · 20% of its score
Beats [Grok 4.20 (0309 reasoning)](grok-4-20.md), [GPT-5.4 mini](gpt-5-4-mini.md), [GPT-5.1](gpt-5-1.md) here, though they rank higher overall. Trails [Llama 3.1 Nemotron Ultra 253B v1](llama-3-1-nemotron-ultra-253b.md), [GLM-4.7](glm-4-7.md), [GPT-5.4 nano](gpt-5-4-nano.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Verified (maker) 70.7% ●····██·████ beats 0/6 hard · avg 21% 6.2% sharp · at its level
SWE-bench Multilingual (maker) 67.7% ●··▄▄···▄▄·█ beats 0/6 hard · avg 30% 4.4% at its level
LiveCodeBench v6 (window unspecif… 89.0% ██···█··●███ beats 3/6 medium · avg 46% 2.8% typical
SciCode subproblems (maker) 44.6% ▃····●··█·▃▃ beats 1/6 hard · avg 19% 2.4% at its level
SciCode (level unspecified) 39.9% ▄·····●█▄▄·▄ beats 1/6 hard · avg 16% 2.2% at its level
LiveCodeBench v6 (Feb–May 2025) 76.4% ▄··●····▄▄·█ beats 2/6 medium · avg 47% 2.1% typical
chat & writing · #84 of 402 · 4.7% of its score
Beats [Claude Haiku 5.5](claude-haiku-5-5.md), [LongCat 2.0 (1.6T-A48B)](longcat-2-0.md), [Muse Glimmer 30B](muse-glimmer-30b.md) here, though they rank higher overall. Trails [Gemma 4 12B](gemma-4-12b.md), [Gemma 4 26B-A4B](gemma-4-26b-a4b.md), [Mistral Medium 3.5 (128B)](mistral-medium-3-5.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LMArena (text) 1426.5 ▁▁▁▂▃▃▃▆▆●█▃ beats 283/385 avg 1260 4.7% blunt
math · #113 of 431 · 6.2% of its score
Beats [Claude Sonnet 4.6](claude-sonnet-4-6.md), [MiniMax-M3](minimax-m3.md) here, though they rank higher overall. Trails [gpt-oss-120b](gpt-oss-120b.md), [Qwen3-VL 32B Thinking](qwen3-vl-32b-thinking.md), [Gemini 2.5 Flash](gemini-2-5-flash.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
IMO-AnswerBench (maker) 88.6% ▄····▄·▄●·█▄ beats 3/6 medium · avg 41% 2.3% typical
OTIS mock AIME 2024–2025 86.7% ▃▃▄▃▃▃▇▇●▃▂▅ beats 135/197 very hard · avg 7.2% 2.0% sharp · noisy
AIME 2026 85.4% ●·▄·▄▄··█··▄ beats 0/6 medium · avg 57% 1.9% sharp · maxed out · noisy
web search · #75 of 84 · 4.4% of its score
Trails [GPT-4.1 nano](gpt-4-1-nano-2025-04-14.md), [GPT-4o mini (2024-07-18)](gpt-4o-mini-2024-07-18.md), [Mistral Small 3.1 24B (2503)](mistral-small-3-1.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
Humanity's Last Exam with tools (… 37.4% ●····▄·█·▄·▄ beats 0/5 hard · avg 15% 2.4% at its level
BrowseComp (maker) 44.4% ▄·●▄···▄··█▄ beats 1/6 hard · avg 18% 2.0% at its level
other tests (15)
test result vs the others in its run difficulty counts why
GPQA Diamond 85.4% ▁▁▁▁▂▄▆▄▅●█▅ beats 147/211 hard · avg 44% 14% sharp · several runs
Humanity's Last Exam (maker) 26.7% ▄··●▄··▄·▄·▄ beats 1/6 very hard · avg 5.7% 8.3% at its level · several runs
GPQA (subset unspecified) (maker) 87.0% ▄●▄▄·▄·····▄ beats 2/6 hard · avg 47% 7.7% sharp
MMLU-Pro (maker) 86.8% ▄·····▄▄●·▄█ beats 3/6 medium · avg 53% 4.5% typical
Global-MMLU-Lite 90.9% ▃······▅··●▃ beats 5/6 medium · avg 61% 3.9% typical
Mystery Game Puzzles (Epoch) 20.0% █··▂▃●▄█▂▄▂▁ beats 35/77 very hard · avg 9.4% 3.0% sharp · too hard for it · noisy
Chess Puzzles (Epoch) 12.0% █▁▁▁▁●▃▃▃▂▁▁ beats 62/137 very hard · avg 5.8% 2.5% sharp · noisy
LongBench v2 (maker) 61.9% ▂·▁·▂▃▁▄▄▃█● beats 29/33 very hard · avg 34% 2.4% at its level
MMLU-ProX (maker) 83.0% ▄·····●·▄▄█▄ beats 1/6 medium · avg 58% 2.3% typical
IFBench (aggregation unspecified) 81.4% ▃······█▃▃·● beats 6/6 medium · avg 45% 2.3% typical
IFBench prompt-level (loose) (mak… 81.7% ▄▄·▄··▄▄···● beats 5/6 medium · avg 47% 2.0% typical
MultiChallenge (maker) 63.8% ▁··········● beats 4/6 hard · avg 28% 1.8% at its level
CritPt (maker) 3.1% ▄····▄●···█▄ beats 2/6 very hard · avg 1.1% 0.9% too hard for it · noisy
WMT25 Terminology Translation (te… 84.8% ▂·▂··▂·●▂·█▆ beats 4/12 medium · avg 63% 0.7% noisy
WMT26 General MT (GEMBA-ESA, LLM … 67.1% ▄·●▄▄·▄██▄█▄ beats 1/12 medium · avg 65% 0.3% blunt
source results share of the score
NVIDIA (submitted / maker-reported) 18 56% ▇▇▇▇▇▇▇
A.X K2 Technical Report 7 25% ▇▇▇
Epoch AI 4 13% ▇▇
LMArena leaderboard dataset 1 5% ▇
Cohere 2 1% ▇
prior: pull toward the middle 0% ▇
weighted average = 68.3
Licences: NVIDIA ([CC BY 4.0](../../licenses/report-2606.15007v1.txt)) · A.X K2 Technical Report ([CC BY 4.0](http://creativecommons.org/licenses/by/4.0/)) · Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Cohere ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/))
## How to run it
### Through an API
api (Together AI) · $1.35/M blended · $0.6 in / $3.6 out per 1M tokens · [docs.together.ai](https://docs.together.ai/docs/serverless-models)
[price source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/providers/togetherai/models/nvidia/nemotron-3-ultra-550b-a55b.toml): Standard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/nvidia/nemotron-3-ultra-550b-a55b.toml)
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Nemotron 3 Ultra 550B-A55B (ud-q4_k_xl, needs ~330.5 GB, from [huggingface.co/unsloth](https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF/tree/main))
```sh
llama cli -hf unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF:UD-Q4_K_XL
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF:UD-Q4_K_XL
```
Other versions: ud-q2_k_xl (~165.5 GB) `llama cli -hf unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF:UD-Q2_K_XL` · ud-iq1_m (~83 GB) `llama cli -hf unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF:UD-IQ1_M`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Nemotron 3 Ultra 550B-A55B (ud-q4_k_xl, needs ~330.5 GB, from [huggingface.co/unsloth](https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF/tree/main))
```sh
lms get https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF@UD-Q4_K_XL -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: ud-q2_k_xl (~165.5 GB) `lms get https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF@UD-Q2_K_XL -y` · ud-iq1_m (~83 GB) `lms get https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF@UD-IQ1_M -y` · 4bit (~309.3 GB, Mac only) `lms get https://huggingface.co/pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-4bit -y` · 8bit (~583.8 GB, Mac only) `lms get https://huggingface.co/pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-8bit -y`
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Nemotron 3 Ultra 550B-A55B (4bit, needs ~309.3 GB, Mac only, from [huggingface.co/pipenetwork](https://huggingface.co/pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-4bit/tree/d64007dfeb7eb87eba579f34fdb03cf329077fd8))
```sh
mlx_lm.chat --model pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-4bit
```
Other versions: 8bit (~583.8 GB, Mac only) `mlx_lm.chat --model pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-8bit`
Model repository and documentation: [huggingface.co/nvidia](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
[interactive version](https://whatmodel.sh/model/nemotron-3-ultra-550b-a55b) · [JSON](https://whatmodel.sh/api?format=json&all=1)