Plain-English Summary
NVIDIA’s very large open reasoning model for enterprise agents, coding, and high-end self-hosted inference.
Nemotron 3 Ultra
NVIDIA’s very large open reasoning model for enterprise agents, coding, and high-end self-hosted inference.
Nemotron 3
Reasoning LLM
GA / open weights
Open weights: Yes
API: NVIDIA NIM and self-hosted
Technical Notes
550B-total, up to 55B-active hybrid Nemotron-H MoE using Mamba and Transformer components, with tool-oriented reasoning.
Best For
Enterprise agents, advanced reasoning, coding, synthetic data
Watch Out For
Requires substantial NVIDIA infrastructure; text-only and complex to deploy.
Model Specs
- Maker
- NVIDIA
- Country
- USA
- Family
- Nemotron 3
- Type
- Reasoning LLM
- Release date
- 2026-07-10
- Date confidence
- Documentation date
- Status
- GA / open weights
- Context window
- Varies
- Max output
- Not listed
- Parameters
- 550B
- Active parameters
- Up to 55B
- Architecture
- Hybrid Mamba-Transformer MoE (Nemotron-H)
- Reasoning
- Yes
- Tool calling
- Yes
- Structured output
- Yes
- API available
- NVIDIA NIM and self-hosted
- Open weights
- Yes
- Self-hostable
- Yes
- License
- NVIDIA Open Model License / model terms
- Inputs
- Text
- Outputs
- Text
- Approx. price
- See provider pricing
- How to access
- Provider website
- Model / API ID
- nvidia-nemotron-3-ultra
- Knowledge cutoff
- Not publicly disclosed
- Fine-tuning
- NIM deployment, fine-tuning, self-hosting
- Completeness
- 90 %
- Verified
- Jul 18, 2026
- Snapshot date
- Jul 18, 2026
- Source
- Official source