NVIDIA · Nemotron 3
Nemotron 3 Ultra 550B A55B
NVIDIA's open-weight (weights, data and recipes) hybrid Mamba-Transformer LatentMoE flagship for agentic, reasoning and long-context workloads.
- AA Intelligence
- 22.9#70 of 151
- Input / output per 1M
- $0.60/ $2.50
- Output speed
- 140tok/s
- Context window
- 1M tokens
Composite indexes
- Artificial Analysis Intelligence Index#70 of 15123
Leader: Claude Opus 5.5 at 58
Coding
- SciCode#74 of 10044.6%40.3%
Leader: Claude Opus 5.5 at 66.9%
- Terminal-Bench 4.0#63 of 960.5%
Leader: Claude Sonnet 5.5 at 70.6%
- Terminal-Bench 2.x#58 of 10556.4%53.9%
Leader: Claude Fable 5.1 at 91.4%
- SWE-bench Verified#10 of 1870.7%
Leader: Claude Sonnet 5 at 85.2%
- LiveCodeBench#4 of 5789%
Leader: Fugu at 92.9%
Agents and tool use
- τ²-Bench#31 of 8983.3%
Leader: GLM-5.2 at 99.1%
- GDPval-AA#58 of 1091000
Leader: Muse Wonder Full at 1857
- BrowseComp#12 of 1644.4%
Leader: GPT-6 Astra at 91.5%
Reasoning and knowledge
- GPQA Diamond#52 of 14487%86.7%
Leader: GPT-6 Astra at 96.1%
- Humanity’s Last Exam#63 of 15337.4%28.4%26.7%
Leader: Claude Sonnet 5.5 at 64.5%
- CritPt#62 of 1443.1%3.1%
Leader: Muse Wonder Full at 33%
- MMLU-Pro#1 of 1586.8%
Leader: Nemotron 3 Ultra 550B A55B at 86.8%
Vision and long context
- AA Long Context Reasoning#44 of 14479.3%65.4%
Leader: Kimi K3 at 88.7%