NVIDIA · Released Mar 5, 2026
Nemotron 3 Ultra
NVIDIA's 550B parameter flagship open architecture designed for high-concurrency enterprise inference clusters.
OpenRating Index
38.6
0–100 composite
Reasoning Index
56.8
0–100 composite
Cost / Task
$0.38
benchmark run
Output speed
141 tok/s
blended tok/s
Time to first token
0.4s
median latency
Input price
$0.2
per 1M tokens
Output price
$0.8
per 1M tokens
Context window
256K
tokens
Strengths
141 tok/s high-throughput generation on GPU clusters
Open weights for internal enterprise deployment
Extensive synthetic data curation and alignment
How we rate Nemotron 3 Ultra
OpenRating scores every model on the same axes. The OpenRating Index is a weighted composite of reasoning, knowledge, and agentic benchmarks, normalised to a 0–100 scale. Speed is blended output measured in tokens per second under standardised load, with time-to-first-token reported separately. Prices are list input and output prices per million tokens, and Task Cost measures end-to-end workload spending.
OpenRating Index
38.6
Reasoning Index
56.8
Token Pricing
Input per 1M tokens
$0.2
Output per 1M tokens
$0.8
Cost of 1M input + 1M output
$1
Task Cost Economics
Empirical cost breakdown for a standard OpenRating Index benchmark task ($0.38 total).
Related models
Compare Nemotron 3 Ultra with the models it is most often measured against.