Text and chat · NVIDIA

Nemotron

NVIDIA's open models for agent work: Super VL reads images and video, Ultra handles long tool-driven tasks, Nano is small and cheap. Private mode.

Private Low cost
Latest versionNemotron 3.5 Super VL 120B A12BOctober 15, 2026
Artificial Analysis Intelligence Index26 points#52 of 90 on Artificial Analysis Intelligence Index
Price$0.562 per million tokens written$0.000393 a page written
Key spec262K tokens of context

What sets it apart

Nemotron 3 is NVIDIA’s open family built on a hybrid Mamba and Transformer mixture-of-experts design, tuned with reinforcement learning on multi-step tool tasks. Ultra scales it to 550B parameters, Super VL adds image and video reading with 12B active, and Nano keeps 3B active for speed and price.

  • Hybrid Mamba, Transformer and MoE design
  • Trained with RL on tool and code tasks
  • Ultra: 550B total, 55B active, open weights

Strengths and limits

Where it shines

  • Ultra is built for long agent runs with many tool calls
  • Private mode, moderate price
  • Super VL reads images and video, has open weights and scores highest of the family on the index

Think twice when

  • Ultra and Nano read text only; images and video come with Super VL
  • Ultra: 550B parameters, 55B active, 256K context here. Super VL: 120B, 12B active, 262K here
  • Open weights

Use cases

Long multi-step agent task

Audit this repository: list the five modules with the most duplicated logic, open each one, and propose a refactoring order with risks.

Ultra is trained for long tool-driven runs.

Cheap classification at scale

Classify each of these 500 support messages as billing, bug, feature request or other, and return one label per line.

Nano is small, fast and among the cheapest models in the catalog.

Reading a screenshot or a scanned page

Here is a screenshot of an error dialog and a photo of an invoice: tell me what the error means and list the invoice lines with their totals.

Super VL reads images, documents and GUIs, which Ultra and Nano do not.

Against its rivals

  • GPT OSSgpt-oss is the other large open reasoning model, with a lower price at similar size.#79 of 90 on Artificial Analysis Intelligence Index (12 points)
  • Qwen open modelsQwen open models also read images; in Nemotron only Super VL does, and Qwen has more sizes to choose from.#23 of 90 on Artificial Analysis Intelligence Index (40 points)
  • GLMGLM scores higher on agent work and costs more.#13 of 90 on Artificial Analysis Intelligence Index (45 points)
Compare side by side ↗

Version history

Each version, dated, with what it changed over the one before.

Artificial Analysis Intelligence Index, version after versionHigher is better · Artificial Analysis

Hollow points are earlier versions no longer offered in Sub Rosa.

See the numbers
VersionReleasedpoints
Nemotron 3 Nano 30B A3B2025-12-157
Nemotron 3 Ultra 550B A55B2026-06-0423
Nemotron 3.5 Super VL 120B A12B2026-10-1526
  1. Nemotron 3 Nano 30B A3B
    • Introduced the family: hybrid Mamba, Transformer and MoE with about 3.5B active parameters
    • Up to 1M token context and a switchable reasoning mode
    Source : huggingface.co
  2. Nemotron 3 Ultra 550B A55B
    • 550B total and 55B active parameters, the largest of the family
    • Intelligence Index 23 versus 7 for Nano, with reasoning and 262K context
    Source : artificialanalysis.ai
  3. Nemotron 3.5 Super VL 120B A12BLatest
    • Vision-language model with 120B total and 12B active parameters, reading text, images and video, with an optional reasoning mode.
    • Scores 26 on the Artificial Analysis index, against 23 for Ultra; open weights under OpenMDW-1.1.
    Source : artificialanalysis.ai

Benchmarks

How to read these scores: what a benchmark tells you.

Artificial Analysis Intelligence Index: #52 of 90 models measured in the catalog. One score averaging about ten independent tests of reasoning, knowledge, math, coding and agent work, all run the same way for every model. Higher is better. Each model is listed at its highest reasoning effort. Two or three points apart is a close call; ten points is a different class of model.

Artificial Analysis Intelligence IndexHigher is better · Artificial Analysis, October 16, 2026
  1. 1. Claude Opus 5.5Claude58
  2. 2. Claude Sonnet 5.5Claude56
  3. 3. Claude Fable 5.1Claude53
  4. 4. GPT-6 AstraGPT53
  5. 5. Gemini 4 ArgonGemini53
  6. 6. GPT-6.1 SolGPT52
  7. 7. Muse WonderMuse52
  8. 8. Claude Opus 5Claude51
  9. 9. Claude Fable 5Claude50
  10. 10. GPT-6 SolGPT48
  11. 52. NVIDIA Nemotron 3.5 Super VL 120BNemotron26
  12. 63. NVIDIA Nemotron 3 UltraNemotron23
  13. 89. NVIDIA Nemotron 3 Nano 30BNemotron7
See the numbers
Modelpoints
1. Claude Opus 5.558
2. Claude Sonnet 5.556
3. Claude Fable 5.153
4. GPT-6 Astra53
5. Gemini 4 Argon53
6. GPT-6.1 Sol52
7. Muse Wonder52
8. Claude Opus 551
9. Claude Fable 550
10. GPT-6 Sol48
52. NVIDIA Nemotron 3.5 Super VL 120B26
63. NVIDIA Nemotron 3 Ultra23
89. NVIDIA Nemotron 3 Nano 30B7

Announced by the maker

Measured by the maker with its own settings: useful, but not directly comparable with the independent scores above.

BenchmarkModelScoreSource
swe-bench-verifiedNVIDIA Nemotron 3.5 Super VL 120B71.1huggingface.co
gpqa-diamondNVIDIA Nemotron 3.5 Super VL 120B89.52huggingface.co

All 3 models in this family

ModelUsePrivacyPrice
NVIDIA Nemotron 3 Ultranvidia-nemotron-3-ultra-550b-a55bTextPrivate$1.405 per million tokens written$0.000984 a page written
NVIDIA Nemotron 3 Nano 30Bnvidia-nemotron-3-nano-30b-a3bTextPrivate$0.135 per million tokens written$0.0000944 a page written
NVIDIA Nemotron 3.5 Super VL 120Bnvidia-nemotron-3-5-super-vl-120b-a12bTextPrivate$0.562 per million tokens written$0.000393 a page written

Specifications

SpecificationNemotron 3.5 Super VLNemotron 3 UltraNemotron 3 Nano
ReleasedOctober 15, 2026June 4, 2026December 15, 2025
Parameters120B total, 12B active550B total, 55B active30B total, 3.5B active
WeightsOpen, OpenMDW-1.1Open, OpenMDWOpen, NVIDIA Nemotron Open Model License
Acceptstext, images, videotexttext
Producestexttexttext
Context262K tokens256K tokens128K tokens
Longest answer·33K tokens16K tokens
LanguagesEnglish plus 6 languages emphasized in post-training·6 languages
AbilitiesReads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for codeUses Sub Rosa’s tools, Thinks before answeringUses Sub Rosa’s tools
Price$0.562 per million tokens written$1.405 per million tokens written$0.135 per million tokens written
PrivacyPrivatePrivatePrivate

Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.