NVIDIA Nemotron 3 Super 120B A12B
NVIDIA Nemotron 3 Super 120B A12B is NVIDIA's 120B total, 12B active-parameter hybrid Mamba-Transformer MoE built for complex multi-agent applications, featuring latent MoE and multi-token prediction.
import { streamText } from 'ai'
const result = streamText({ model: 'nvidia/nemotron-3-super-120b-a12b', prompt: 'Why is the sky blue?'})Playground
Try out NVIDIA Nemotron 3 Super 120B A12B by NVIDIA. Usage is billed to your team at API rates. Free users (those who haven't made a payment) get $5 of credits every 30 days.
About NVIDIA Nemotron 3 Super 120B A12B
NVIDIA released NVIDIA Nemotron 3 Super 120B A12B on March 18, 2026 as the second model in the Nemotron 3 family, following Nano. It has 120B total parameters and 12B active parameters per token. The hybrid Mamba-Transformer MoE backbone interleaves Mamba-2 layers for long-sequence processing, Transformer attention layers for precise recall, and MoE layers for compute efficiency. NVIDIA Nemotron 3 Super 120B A12B delivers higher throughput than the previous Nemotron Super generation.
Two architectural innovations distinguish Super from Nano. First, latent MoE: before routing, token embeddings compress into a low-rank latent space. This lets the model consult 4x as many expert specialists at the same inference cost. Finer-grained routing allows distinct experts to activate for different subtasks (Python syntax, SQL logic, multi-hop reasoning) without paying the compute cost of running them all. Second, multi-token prediction (MTP): the model predicts multiple future tokens in a single forward pass. MTP strengthens reasoning during training and provides built-in speculative decoding at inference, yielding up to 3x speedups on structured generation tasks like code and tool calls.
On PinchBench (a benchmark evaluating LLMs as the planning brain of an OpenClaw agent), NVIDIA Nemotron 3 Super 120B A12B scores 85.6%. Full announcement: https://docs.aws.amazon.com/en_us/bedrock/latest/userguide/model-card-nvidia-nemotron-super-3-120b.html.
Providers
Route requests across multiple providers. Copy a provider slug to set your preference. Visit the docs for more info. Using a provider means you agree to their terms, listed under Legal.
P50 throughput on live AI Gateway traffic, in tokens per second (TPS). Visit the docs for more info.
P50 time to first token (TTFT) on live AI Gateway traffic, in milliseconds. View the docs for more info.
Direct request success rate on AI Gateway and per-provider. Visit the docs for more info.
More models by NVIDIA
| Model |
|---|
| 131K | 0.2s | 143tps | $0.06/M | $0.23/M | — | — |   | 08/18/2025 | |||||
| 262K | 0.2s | 601tps | $0.05/M | $0.24/M | — | — | | 12/01/2024 | |||||
| 131K | 0.2s | $0.20/M | $0.60/M | — | — | | 12/01/2024 |
What To Consider When Choosing a Provider
- Configuration: NVIDIA Nemotron 3 Super 120B A12B's multi-agent orientation means it works best as the planning and reasoning backbone in a pipeline where lighter models handle individual steps. Evaluate your task decomposition before choosing a tier. Compare $0.15 and $0.65.
- Zero Data Retention: AI Gateway supports Zero Data Retention for this model via direct gateway requests (BYOK is not included). To configure this, check the documentation.
- Authentication: AI Gateway authenticates requests using an API key or OIDC token. You do not need to manage provider credentials directly.
When to Use NVIDIA Nemotron 3 Super 120B A12B
Best For
- Complex multi-agent applications: Software development pipelines or cybersecurity triaging that require deep planning across long contexts
- Context explosion workloads: Multi-agent systems with up to 15x the token volume of standard chats that cause goal drift with smaller models
- Dense technical problem-solving: Tasks where higher parameter count provides reasoning headroom
- Super plus nano pattern: Agentic pipelines pairing Super for complex decisions with Nano for efficient individual steps
- Fully open model requirement: Teams that need weights and recipes for enterprise customization, data control, or reproducibility
Consider Alternatives When
- Simpler task steps: Nemotron 3 Nano is more throughput-efficient for lighter workloads
- Vision-language inputs: Super is text-only; Nemotron Nano 12B v2 VL supports multimodal inputs
- Cost-first constraints: A lighter model may deliver acceptable quality at lower cost per token
Conclusion
NVIDIA Nemotron 3 Super 120B A12B combines latent MoE for expert specialization and multi-token prediction for inference speedups. Route requests through AI Gateway as the planning and reasoning backbone for complex multi-agent applications at scale.
Frequently Asked Questions
What is latent MoE and why does it matter?
Latent MoE compresses token embeddings into a smaller latent space before routing. This reduces per-expert compute cost and lets NVIDIA Nemotron 3 Super 120B A12B consult 4x as many experts for the same inference budget. Distinct experts activate for code generation, SQL logic, and natural language without the overhead of running them all densely.
What is multi-token prediction and how does it speed up inference?
MTP trains the model to predict multiple future tokens in a single forward pass. At inference, the MTP heads provide draft tokens that can be verified in parallel, acting as built-in speculative decoding. This delivers wall-clock speedups for structured generation like code and tool calls, without requiring a separate draft model.
What is the "Super + Nano" deployment pattern?
NVIDIA describes using Nano for straightforward individual steps in a pipeline and Super for complex decisions requiring deep reasoning. In software development, for example, Nano might handle routine merge requests while Super tackles tasks that require understanding a full codebase. This pattern distributes compute across task difficulty.
What is the context window of 256K tokens used for in multi-agent systems?
Multi-agent systems generate high token volume (up to 15x that of standard chats) from tool outputs, reasoning steps, and history resent at each turn. A window of 256K tokens lets agents keep full session history, large codebases, and retrieved context in a single pass. This reduces goal drift from context truncation.
Where are hosted per-token prices?
Rates are listed on this page. They reflect the providers routing through AI Gateway and shift when providers update their pricing.