Nemotron 3 Ultra

NVIDIA's fully open (weights, data and recipes) flagship, released June 4, 2026 — the leading US open-weights model on Artificial Analysis' Intelligence and Agentic indexes. A hybrid Mamba-Attention LatentMoE (550B/55B active) with 400+ tok/s throughput, SOTA open-model long-context retrieval at 1M tokens, and inference-time reasoning budget control. Free to self-host; hosted via OpenRouter/NIM/SageMaker.

Nemotron 3 Ultra strengths

• Leading US open-weights on AA indexes
• 400+ tok/s throughput
• SOTA open-model 1M-context retrieval
• Fully open: weights, data, recipes
• Reasoning budget control

Pricing & context

Cost guide: a typical call of about 10K input + 2K output tokens costs roughly $0.009 at list prices. Worth modelling against cheaper tiers before committing high-volume traffic.

When to choose Nemotron 3 Ultra

Nemotron 3 Ultra is best for long-running agent workflows, long-context processing and cost-efficient self-hosted deployments. If your workload is more cost-sensitive, weigh it against gpt-oss-120b (≈$0.03 input /1M) first.

Nemotron 3 Ultra FAQ

How much does Nemotron 3 Ultra cost?

Nemotron 3 Ultra is priced at $0.50 per 1M input tokens and $2.20 per 1M output tokens (public API list price), with a 1M tokens context window. A typical call of about 10K input and 2K output tokens costs roughly $0.009.

What is Nemotron 3 Ultra best for?

Nemotron 3 Ultra by NVIDIA is best for long-running agent workflows, long-context processing and cost-efficient self-hosted deployments.

How does Nemotron 3 Ultra pricing compare to Step 3.7 Flash?

Nemotron 3 Ultra input costs $0.50 per 1M tokens versus $0.20 for Step 3.7 Flash, roughly 2.5x more expensive on input. Output is $2.20 vs $1.15.

Is Nemotron 3 Ultra multimodal?

Nemotron 3 Ultra supports text.