NVIDIA: Nemotron 3 Ultra

NVIDIA

open_source_flagship

Capabilities

ToolsStructured OutputExtended ThinkingLong ContextCode

Strengths

Frontier reasoning and long-context (512K tokens)
Open weights enable self-hosting and data privacy
Strong extract and generation across benchmarks
Efficient MoE inference relative to parameter count

Weaknesses

Very slow inference (0.2 speed_score; 211 tok/s)
Weaker code performance vs other 500B+ models
Training data provenance not fully disclosed
No vision or multimodal capabilities

Pricing

Input / 1M tokens

$0.5

Output / 1M tokens

$2.2

Context window

512.288k

Transparency

Open weights

10.0 / 10

Open training data

0.0 / 10

Open methodology

7.5 / 10

Licence openness

8.0 / 10

Provider disclosure

7.0 / 10

FMTI company score

N/A

Composite:6.5 / 10

Open-weight 550B MoE model (55B active) with Apache 2.0 licence. NVIDIA publishes architecture and training approach, but training data composition not fully disclosed. Can be self-hosted for full data privacy.

Sustainability

Inference energy

6.5 / 10

Training footprint

N/A

Provider infrastructure

7.2 / 10

Composite:6.8 / 10

MoE architecture with sparse activation reduces per-inference compute vs dense 550B. NVIDIA has GPU efficiency investments. Training footprint unknown but large-scale pretraining implies substantial baseline.

Task Fitness

question answering

71%

code

74%

mathematics

88%

business communication

81%

vision

analysis

73%

extraction

89%

long-form generation

86%

research

80%

reasoning

71%

summarisation

85%

translation

78%

conversation

81%