What Nemotron 3 Ultra delivers

Nemotron 3 Ultra scores 71.9% on SWE-bench Verified, 87% on MMLU-Pro, and achieves the highest Artificial Analysis Intelligence Index (48) among US-developed open-weight models. It maintains 91.75% on RULER at 1M tokens — far outperforming competitors on long-context retention.

At 300+ tokens per second on commodity US GPU infrastructure, it is also one of the fastest models for its capability class.

The Nemotron 3 family

Nano: Smallest variant for edge and consumer GPU deployment.

Super: Mid-range with strong long-context retention (91.75% at 1M tokens on RULER).

Ultra: Flagship with 71.9% SWE-bench Verified, best for enterprise coding and reasoning.

Deployment options

Self-host: Download weights from Hugging Face. Optimized for NVIDIA GPUs with TensorRT-LLM.

API: Free on OpenRouter with pay-per-token pricing from partners.

Cloud: Available on major cloud GPU providers with pre-configured NVIDIA NIM containers.

Best for US enterprise deployments

Data residency requirements make Nemotron 3 Ultra the safest choice for US-based organizations that cannot use Chinese-origin model weights.

For pure intelligence score without compliance constraints, Kimi K2.6 and GLM 5.2 score higher overall.

Ultra's throughput advantage (300+ tok/s) makes it ideal for high-volume agent pipelines.

Training infrastructure is also open

NVIDIA released the entire training infrastructure: NeMo Data Designer for synthetic data, the RL training pipeline, and the NeMo ecosystem tools.

This makes Nemotron 3 one of the most reproducible open-weight model releases.

FAQ

Can Nemotron 3 Ultra run on consumer hardware?

Ultra requires multi-GPU setups. The Nano variant is designed for consumer GPUs and edge deployment.

What license does Nemotron 3 use?

NVIDIA's custom open-weight license. Review terms for commercial use restrictions on the Hugging Face model card.

Deploy Nemotron 3 Ultra

  1. Choose variant: Ultra for flagship, Super for mid-range, Nano for edge
  2. Download weights from Hugging Face
  3. Serve with TensorRT-LLM for maximum throughput on NVIDIA GPUs
  4. Configure with 1M context for long-document workloads
  5. Deploy via NIM container for production
  6. Benchmark throughput against your latency requirements

Ready to see the full catalog?

Browse all 42 agent-ready business kits. Each includes pre-configured agent roles, prompts, workflows, install steps, and approval gates.

Shop all business kits