Model comparison
Compare NVIDIA Nemotron 3 Ultra and NVIDIA Nemotron 3.5 Lightning using the same provider-sourced text & reasoning rubric. No mystery score and no invented benchmark ranking.
Facts checked September 4, 2026
Set your usage. Your estimate updates as you type.
Example: 6,000 input tokens for 10 pages, plus a 500-token summary. Page lengths vary; adjust the numbers below.
One run sends your input to the model once and receives an answer. Tokens are pieces of text: input is what you send, output is the answer you receive.
Estimates exclude taxes, tools, cache storage/writes, free allowances and custom discounts. Image estimates cover output only, not prompt or reference-image charges. Quality modes differ by model. Unlisted settings are not treated as free.
| Model | Access | Estimated total (USD) |
|---|---|---|
| NVIDIA Nemotron 3 UltraNVIDIA | NVIDIA | No reviewed rate |
| NVIDIA Nemotron 3.5 LightningNVIDIA | NVIDIA | No reviewed rate |
Quick take
NVIDIA's largest Nemotron 3 reasoning model for complex agents, coding, planning, tools, RAG, and million-token analysis.
Self-hosting Ultra is a substantial hardware project. Size the real context cache, concurrency, precision, and availability requirements before treating open weights as the cheaper route.
NVIDIA's compact 30B mixture-of-experts model for efficient specialist agents and high-volume text workflows.
The current release is labeled preview. Validate the exact precision, language, tool template, provider route, and long-context memory needs before standardizing a production fleet.
Compare the published facts
Values use each provider's own published units and limits. A blank means the provider did not publish a directly comparable value in the sources reviewed.
| Text & reasoning | NVIDIA Nemotron 3 Ultra | NVIDIA Nemotron 3.5 Lightning |
|---|---|---|
| Context windowMaximum combined prompt and working context documented by the provider. | Up to 1M tokens | Up to 1M tokens |
| Maximum outputProvider-published response limit, where available. | Not separately published | Not separately published |
| Knowledge cutoffLatest reliable knowledge date explicitly published by the model provider. Search and connected tools can retrieve newer information but do not change the model's built-in cutoff. | Pretraining through Sep 2025; post-training through May 2026 | Pretraining through Sep 2025; post-training through May 2026 |
| Input priceCurrent standard list price per million input tokens unless noted. | Free prototype or deployment cost | Free prototype or deployment cost |
| Output priceCurrent standard list price per million output tokens unless noted. | Free prototype or deployment cost | Free prototype or deployment cost |
| InputsMedia types accepted by the listed model endpoint. | Text | Text |
| Tools & agentsSelected native tools and agent-building capabilities, not an exhaustive list. | Tool calling, agents, configurable reasoning budget | Agentic tools and long-running workflows |
How to choose
Start with the job you need to complete, then validate cost, access, and policy details on your exact provider route.
Downloaded weights can run in infrastructure the operator controls. NVIDIA's hosted prototype and partner endpoints are governed by their own trial, logging, retention, and deployment terms.
Provider and API links
Self-hosted weights keep request data under the operator's controls. NVIDIA API trials, OpenRouter, and cloud partners each apply separate logging, retention, and training-use policies.
Provider and API links
Frequently asked questions
NVIDIA Nemotron 3 Ultra: NVIDIA's largest Nemotron 3 reasoning model for complex agents, coding, planning, tools, RAG, and million-token analysis. NVIDIA Nemotron 3.5 Lightning: NVIDIA's compact 30B mixture-of-experts model for efficient specialist agents and high-volume text workflows.
Consider NVIDIA Nemotron 3 Ultra when your priority is Complex long-running agents. Consider NVIDIA Nemotron 3.5 Lightning when your priority is High-volume agent sub-tasks. Test both with your own data and provider route before committing.
No. This comparison aligns provider-published facts for the Text & reasoning category. It does not claim a universal winner or combine incompatible third-party benchmark scores.
