Nemotron 3 Ultra
NVIDIA · released
NVIDIA's largest open-weights model: a 550-billion-parameter hybrid design for long-running agents, coding and long-context analysis. NVIDIA releases the weights, much of the training data and the recipes rather than selling API access.
Best for: Self-hosting on NVIDIA hardware · Customising with open data and recipes · Fast output
- Intelligence Index
- 22.9
- #31 of 38 · at Reasoning effort
- Price per 1M tokens
- $0.60 in $2.5 out
- Blended $1.07 at 3:1
- Context window
- 1M
- Output speed
- 155 tokens/s
- $0.60 measured cost per task
Capability profile
One 0–100 score per area, placing this model between the weakest and strongest tracked model across that area's benchmarks.
- Overall33 · #29 of 38
Based on 2 of 3 benchmarks
- Coding7 · #35 of 39
Based on 2 of 5 benchmarks
- Agents & knowledge work38 · #27 of 40
Based on 1 of 6 benchmarks
- Reasoning & science23 · #31 of 39
Based on 2 of 2 benchmarks
- Knowledge & reliability62 · #24 of 39
Based on 2 of 2 benchmarks
- VisionNot enough data
Based on 0 of 2 benchmarks
Benchmark results
All third-party results on record. Rank is among the tracked models that have been tested on that benchmark.
Overall
#31 of 38· Reasoning effort· Artificial Analysis
- Vals IndexNot tested
- Arena · Text1426 ±7
#29 of 33· 11,515 votes· Arena
Coding
#31 of 38· Reasoning effort· Artificial Analysis
- SciCode40.3%
#34 of 38· Reasoning effort· Artificial Analysis
- Arena · WebDevNot tested
- Vibe Code BenchNot tested
- Code MigrationNot tested
Agents & knowledge work
- GDPval-AA25.8
#30 of 39· Reasoning effort· Artificial Analysis
- Arena · AgentNot tested
- Finance Agent v2Not tested
- Excel Modeling BenchmarkNot tested
- Legal Research BenchNot tested
- Tax Agent BenchNot tested
Reasoning & science
- Humanity's Last Exam28.4%
#31 of 39· Reasoning effort· Artificial Analysis
- CritPt3.1%
#31 of 39· Reasoning effort· Artificial Analysis
- ARC-AGI-2saturatingNot tested
- ARC-AGI-3Not tested
- GPQA Diamondsaturating86.7%
#19 of 27· Reasoning effort· Artificial Analysis
Knowledge & reliability
#23 of 39· Reasoning effort· Artificial Analysis
- AA-LCR79.3%
#26 of 39· Reasoning effort· Artificial Analysis
Vision
- MMMU-ProNot tested
- Arena · VisionNot tested
Reported by NVIDIA
Numbers the lab published at launch, run on its own harness and settings. Useful context, but not independently verified and never used in the rankings on this site.
- SWE-bench Verified · NVIDIA-run70.7%
- Terminal-Bench 2.1 · NVIDIA-run56.4%
- GPQA Diamond · no tools87%
- Humanity's Last Exam · no tools26.7%
Specifications
- Lab
- NVIDIA
- Released
- Jun 4, 2026
- Availability
- Generally available
- Context window
- 1,048,576 tokens
- Max output
- —
- Accepts
- Text
- Knowledge cutoff
- Sep 2025
- Weights
- Open
- Licence
- OpenMDW 1.1
- Parameters
- 550B total · 55B active
- Reasoning control
- Thinking on or off
- API model ID
nvidia/nemotron-3-ultra-550b-a55b
Pricing
- Input / 1M tokens
- $0.60
- Cached input / 1M
- $0.16
- Output / 1M tokens
- $2.5
- Blended (3:1)
- $1.07
- Measured cost / task
- $0.60
Typical price at third-party hosts. NVIDIA offers a free trial endpoint but no paid first-party API.
Worth knowing
- The model supports up to 1M tokens of context, but hosted endpoints typically cap it near 262K.
Official sources
Closest alternatives
The models nearest to this one on overall score.
23 benchmarks tracked. See methodology and sources.