FastMetal
Back to Models
nvidia logoNVIDIA
nemotron-3-super-120b-a12b

Nemotron 3 Super

NVIDIA Nemotron 3 Super is a 120B-parameter open hybrid MoE model, activating just 12B parameters for maximum compute efficiency and accuracy in complex multi-agent applications. Built on a hybrid Mamba-Transformer Mixture-of-Experts architecture with multi-token prediction (MTP), it delivers over 50% higher token generation compared to leading open models. The model features a 1M token context window for long-term agent coherence, cross-document reasoning, and multi-step task planning. Latent MoE enables calling 4 experts for the inference cost of only one, improving intelligence and generalization. Multi-environment RL training across 10+ environments delivers leading accuracy on benchmarks including AIME 2025, TerminalBench, and SWE-Bench Verified. Fully open with weights, datasets, and recipes under the NVIDIA Open License, Nemotron 3 Super allows easy customization and secure deployment anywhere — from workstation to cloud.

3/11/2026
262,144 tokens
#165 Text (spanish)
Input: $0.0897/M
Output: $0.422/M
Try in ChatGet API Key

Specifications

Modalities

Input
text
Output
text

Supported Parameters

frequency_penalty
include_reasoning
max_tokens
min_p
presence_penalty
reasoning
repetition_penalty
response_format
seed
stop
temperature
tool_choice
tools
top_k
top_p

Reasoning Configuration

Default
Thinking on (effort: medium)
Selectable effort levels
medium
low
Turning thinking off
Possible (reasoning.enabled: false)

How to control thinking, and what it costs

Data policy

Prompt retention
Unknown — we could not confirm
Training
Not used for training

"Unknown" does not mean "safe". It means we could not confirm it.

Whether a provider trains on prompts is a declared value from our terms with them. Retention is determined from the upstream listing for every host this model can reach. Neither is guessed.

Measured performance

Time to first response (p50)617ms
Throughput (p50)75 tokens/s

    Measured through FastMetal's own gateway over the last 30 days (497 first-response samples).

    Throughput is sampled once a day: the median of 23 observations over the last 30 days.

    Time to first response is measured to the first streamed chunk; throughput counts all output tokens, including reasoning tokens. These are observations, not a performance guarantee.

    Code Examples

    curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "nemotron-3-super-120b-a12b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

    How Nemotron 3 Super actually answers

    Real responses to our standard prompts, recorded on FastMetal.

    Count the number of 'r's in 'strawberry'

    Count the number of 'r's in 'strawberry'. Explain your reasoning step by step.

    ### Step-by-Step Reasoning for Counting 'r's in 'strawberry' 1. **Write out the word clearly**: The word is "strawberry". To avoid confusion, I'll list each letter in sequence with its position: - Position 1: `s` - Position 2: `t` - Position 3: `r` - Position 4: `a` - Position 5: `w` - Position 6: `b` - Position 7: `e` - Position 8: `r` - Position 9: `r` - Position 10: `y` 2. **Examine each letter individually for 'r'**: - Position 1 (`s`): Not 'r' → count remains 0. - Position 2 (`t`): Not 'r' → count remains 0. - Position 3 (`r`): **Is 'r'** → count increases to 1. - Position 4 (`a`): Not 'r' → count remains 1. - Position 5 (`w`…
    66.1s response1861 output tokens

    Debug This Error

    I'm getting the following error in my Node.js application: TypeError: Cannot read properties of undefined (reading 'map') at UserList (/app/components/UserList.js:12:25) at renderWithHooks (/app/node_modules/rea…

    **Why the error occurs** ```text TypeError: Cannot read properties of undefined (reading 'map') ``` `Array.prototype.map` is being called on the value `undefined`. In your component: ```jsx function UserList({ users }) { return ( <ul> {users.map(user => <li key={user.id}>{user.name}</li>)} </ul> ); } ``` `users` is `undefined` when the component renders, so `users.map` throws. Typical reasons: | Situation | What happens | |-----------|--------------| | Parent component forgets to pass the `users` prop (`<UserList />`) | `users` is `undefined` | | Parent passes `users={null}` or the result of an async fetch that hasn’t resolved yet (`users` stays `undefined` until t…
    11.4s response1144 output tokens

    Code Review

    Please review the following Python function and suggest improvements for readability, performance, and best practices: def get_data(url, retries=3): import requests import time for i in range(retries):…

    **Original code – quick recap** ```python def get_data(url, retries=3): import requests import time for i in range(retries): try: r = requests.get(url, timeout=10) if r.status_code == 200: data = r.json() result = [] for item in data: if item["active"] == True: result.append({"name": item["name"], "id": item["id"]}) return result else: time.sleep(2) except: time.sleep(2) return [] ``` Below are concrete, actionable suggestions grouped by **readability**, **performance**, and **best‑practic…
    60.6s response3361 output tokens

    Compare these answers side by side with other models →

    Frequently asked questions

    How much does the Nemotron 3 Super API cost?
    On FastMetal, Nemotron 3 Super is billed per token: $0.0897 per 1M input tokens and $0.422 per 1M output tokens on a US-dollar account, before tax. Usage is drawn from a prepaid balance; there is no subscription or monthly fee.
    Can I call Nemotron 3 Super with the OpenAI SDK?
    Yes. Point base_url at https://api.fastmetal.ai/v1 and pass "nemotron-3-super-120b-a12b" as the model; existing OpenAI-style code works unchanged, including streaming, tool calls and structured output.
    What is the context window of Nemotron 3 Super?
    262,144 tokens, shared between the prompt and the response.
    How does Nemotron 3 Super rank?
    #208 on the public arena's Overall board (ELO 1,360). Ranks move as the leaderboard is updated.
    What do I need to try Nemotron 3 Super?
    Create an account and add credit; the model is then available both in the browser chat and over the API. There is no contract or minimum spend.

    Try Nemotron 3 Super right now

    Nemotron 3 Super is available on FastMetal through one API key. Start in the browser, or call it from the OpenAI SDK.

    Create a free account

    Leaderboard

    Text
    spanishELO: 1,389
    #165
    russianELO: 1,326
    #236
    CodingELO: 1,407
    #208
    MathELO: 1,371
    #200
    Creative WritingELO: 1,303
    #236
    Instruction FollowingELO: 1,342
    #217
    Hard PromptsELO: 1,378
    #208
    Multi-TurnELO: 1,350
    #211
    OverallELO: 1,360
    #208
    ChineseELO: 1,401
    #190
    EnglishELO: 1,384
    #201
    frenchELO: 1,390
    #184
    germanELO: 1,314
    #203

    Rankings: LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0, snapshot 2026-10-22; category names mapped and ratings rounded by FastMetal.

    More from NVIDIA

    Llama 3.3 Nemotron Super 49B V1.5
    131,072 tokens
    Nemotron 3 Nano 30B A3B
    262,144 tokens
    Llama 3.1 Nemotron 70B Instruct
    131,072 tokens
    Llama 3.1 Nemotron Ultra 253B v1
    131,072 tokens
    Nemotron 3.5 Super VL
    262,144 tokens
    Nemotron 3 Ultra
    512,288 tokens