Nemotron 3 Super review: accurate answers, slow responses
Nemotron 3 Super (nemotron-3-super-120b-a12b) is for work where a careful answer matters more than a fast one. It fixed honorific errors correctly and, given a word problem with contradictory numbers, said so instead of inventing an answer. Its weakness is speed: it reasons before answering, and some replies took close to a minute. It fits batch jobs and agent back ends better than interactive chat.
Specs: a hybrid Mamba-Transformer MoE
NVIDIA released Nemotron 3 Super as an open model in March 2026.
- Mixture of experts: 120B total parameters, 12B active
- Hybrid architecture combining Mamba and Transformer layers
- Context: about 260,000 tokens
- Input and output: text
- Tool calls; reasons before it answers
Five writing and reasoning tasks
We sent the same five tasks, all written in Japanese, to several models.
| Task | Result | Response time | Output tokens (reasoning) |
|---|---|---|---|
| Apology email for a late delivery | Polite and natural, though it used a formal letter closing without the matching opening | ~30 s | 1,033 (980) |
| Three-point summary | Concise and accurate | ~4 s | 153 (26) |
| Fix two honorific errors | Correct, with the reasons laid out in a table | ~44 s | 1,963 (1,533) |
| Math word problem | Correctly said there is no solution, and showed the nearest integer combinations | ~57 s | 1,987 (1,272) |
| Haiku with a seasonal word | 秋の風 デバッグ夜半 夜明けかな ("autumn wind; debugging past midnight, and then dawn"). Correct seasonal word | ~69 s | 3,296 |
The word problem was built so that no integer solution exists. Nemotron 3 Super set up the equations, stated that no count satisfies both conditions, and worked out the totals for 7 and 8 items to show how close each comes. A model that will not make up an answer is worth a lot in production.
The flaw was the email. A formal Japanese business letter that ends with the set closing phrase must also open with its paired salutation; Nemotron 3 Super wrote the closing alone, and several models in the group made the same mistake. When you generate formal templates, spell out the opening and closing in the prompt.
Speed and billing
Tasks with a fixed answer, like the summary, came back in seconds. Tasks with room to think, like explaining a correction or writing a haiku, spent 1,000 to over 3,000 tokens reasoning and took 30 seconds to a minute.
- Reasoning tokens are billed as output tokens.
- It suits jobs that read a lot and write a little, such as summarizing or extracting from long documents.
- It is a poor fit for a chat UI where the user is waiting for the reply.
It is billed per million input and output tokens; current prices are on the model page.
Good fit, poor fit
Good fit
- Summarizing and extracting from long documents (large input, small output)
- Tool calls inside an agent, out of the user's sight
- Checks where you want contradictions in numbers or conditions flagged
Poor fit
- Chat that needs an answer within a few seconds
- Formal documents with strict formatting conventions, unless you supply a template
FAQ
Q. Why are Nemotron 3 Super's responses slow?
It reasons before answering. On simple tasks such as summarizing, the reasoning is short and the reply arrives in seconds.
Q. Does Nemotron 3 Super support tool calls?
Yes. Pass tools through FastMetal's OpenAI-compatible API.
Q. How did it handle non-English tasks?
All five tasks were in Japanese. The honorific fixes and the math were correct; formal letter conventions needed help from the prompt.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nemotron-3-super-120b-a12b",
"messages": [{"role": "user", "content": "Summarize in three bullet points: Small businesses are adopting generative AI, but unpredictable costs remain a barrier. Bills in a foreign currency move with exchange rates and can come in higher than planned."}]
}'
The same API key and balance work across the catalog, so you can change model and compare. Every model is in the model catalog.