NVIDIA's hybrid LatentMoE model (120B total, 12B active). Mamba-2 + Attention + MoE architecture with 1M context. Multi-Token Prediction for fast inference.
NVIDIA Nemotron 3 Super (120B total, 12B active) is an open-weight hybrid LatentMoE model built for agentic reasoning, coding, and long-context multi-agent workflows. Released in March 2026, it is the flagship of NVIDIA's Nemotron 3 generation — the first model to use LatentMoE architecture, compressing token representations before expert routing to activate 4× more experts at the same inference cost. The model natively supports a 1M-token context window, giving AI agents long-term memory spans across codebases, document corpora, and extended multi-turn tool-using sessions.
The model interleaves Mamba-2 state-space layers for efficient long-sequence processing with attention layers for precise associative recall, and uses Multi-Token Prediction (MTP) to forecast multiple future tokens per forward pass — enabling built-in speculative decoding. Pre-trained for over 25T tokens and fine-tuned via multi-environment reinforcement learning (1.2M+ NeMo Gym rollouts), it excels in agentic benchmarks like PinchBench (85.6%), SWE-Bench verified, and Terminal-Bench while offering up to 5× throughput over its predecessor. On TheRouter, it is routed via AWS Bedrock as nvidia.nemotron-super-3-120b.
NVIDIA publishes Nemotron 3 Super as open weights, datasets, recipes, NIM deployment, and a native 1M-token context model. AWS Bedrock serves the same model as nvidia.nemotron-super-3-120b with 256K context, 32K max output, text-only I/O, Chat Completions, Invoke, Converse, response streaming, structured outputs, guardrails, and client-side tool calling.
TheRouter exposes it as OpenAI-compatible /v1/chat/completions under nvidia/nemotron-super-120b. TheRouter's live catalog supplies the customer-facing context, output, pricing, modality, and supported-parameter fields; this page records that the upstream Bedrock card is narrower than NVIDIA's native 1M context and that snippets remain capped until operator-funded verification runs.
| Type | Rate |
|---|---|
| Input | $0.162 / 1M tokens |
| Output | $0.702 / 1M tokens |
| Release date | 2026-03-11developer.nvidia.com ↗ | verified |
| Architecture | Hybrid LatentMoE — Mamba-2 SSM + Attention + MoE (120B total, 12B active)research.nvidia.com ↗ | verified |
| Native context length | 1,000,000 tokens (1M)developer.nvidia.com ↗ | verified |
| Context on AWS Bedrock | 256K tokens (AWS Bedrock exposes a subset of the native 1M)docs.aws.amazon.com ↗ | verified |
| Max output (TheRouter) | 262,144 tokensdocs.aws.amazon.com ↗ | verified |
| Training data | 25T+ tokens — crawled and synthetic code, math, science, general knowledge; datasets fully disclosedbuild.nvidia.com ↗ | verified |
| Post-training | Multi-environment RL (1.2M+ NeMo Gym rollouts, 21 environments)developer.nvidia.com ↗ | verified |
| License | NVIDIA Open Model Licensewww.nvidia.com ↗ | verified |
| Supported languages | English, French, German, Italian, Japanese, Spanish, Chineseaws.amazon.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
PinchBench Agent orchestration benchmark — best open model in class at launch | 85.6%% | developer.nvidia.com ↗ | |
AIME 2025 | — | Not publicly disclosed | — |
MMLU-Pro | 83.73% | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | |
GPQA (with tools) | 82.70%% | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | |
Arena-Hard V2 Lower than dedicated chat models — model optimized for agentic execution, not conversation | 73.88%% | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | |
RULER Long-context retrieval benchmark — 100 samples per task, strong on needle-in-haystack | High (leading open model) | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | |
TerminalBench 2.0 | 31.00points | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | |
SWE-Bench Verified AWS blog highlights strong performance on SWE-Bench and multilingual variants, no exact figure published | — | Not publicly disclosed | — |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "nvidia/nemotron-super-120b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Nemotron 3 Super supports OpenAI-compatible chat completions through TheRouter's API. The model supports tools, tool_choice, response_format (JSON mode), stop sequences, temperature, top_p, and max_tokens.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/nemotron-super-120b",
"messages": [
{"role": "system", "content": "You are a helpful AI assistant."},
{"role": "user", "content": "Write a Python function to recursively find all .py files in a directory."}
],
"temperature": 0.3,
"max_tokens": 4096
}'NVIDIA open-released Nemotron 3 Super (120B-A12B), a hybrid Mamba-2 + Attention + LatentMoE model with 1M native context. Built for multi-agent systems, it features multi-token prediction, NVFP4 pretraining for Blackwell, and multi-environment RL post-training across 21 configurations. Available via AWS Bedrock from day one.
AWS published a detailed guide for running Nemotron 3 Super on Bedrock, covering use cases from software development to cybersecurity triage. The model is available through both bedrock-runtime and bedrock-mantle endpoints with Standard, Priority, and Flex tiers.
DeepInfra announced support for Nemotron 3 Super with optimized inference. The platform highlighted its LatentMoE architecture and 1M context window as key differentiators for agentic AI workloads.
In standard MoE, tokens are routed to experts at the full hidden dimension, which creates a computational bottleneck as the model grows. LatentMoE compresses token embeddings into a low-rank latent space before routing, allowing up to 4× more experts to be consulted at the same inference cost. This enables finer-grained expert specialization — for example, distinct experts for Python syntax vs. SQL logic — that is only activated when needed.
The model's native architecture supports a 1M-token context window, but the exact exposed limit depends on the serving platform. On AWS Bedrock (TheRouter's current routing backend for this model), it is configured for 256K tokens with 32K max output. The model's Mamba-2 layers make the full 1M context computationally practical, unlike attention-only models where long context degrades throughput quadratically.
Nemotron 3 Super is the first model in the Nemotron 3 generation. Its predecessor was the earlier Nemotron Super model (a dense 120B). The 3 Super delivers up to 5× throughput improvement via its hybrid MoE architecture (dense → 12B active), introduces LatentMoE for 4× more expert routing, uses multi-token prediction for speculative decoding, and supports a native 1M context vs. the previous generation's shorter window. A smaller sibling (Nano 30B, 3.5B active) was released in December 2025 for edge deployment.
It works fine for chat, but its Arena-Hard V2 score (73.88%) indicates it lags behind specialist chat models. Nemotron 3 Super is optimised for agentic, multi-turn problem-solving where tool use, long context, and high throughput matter. For open-ended creative conversations, consider a model like deepseek/deepseek-v3.1 or meta/llama-4-maverick instead.
Nemotron 3 Super supports temperature, max_tokens, top_p, tools (function calling), tool_choice, response_format (JSON mode), and stop sequences. Streaming is supported. The model operates via TheRouter's v1/chat/completions OpenAI-compatible endpoint routed through AWS Bedrock.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | developer.nvidia.com ↗ | 2026-05-28 | verified |
| Architecture | research.nvidia.com ↗ | 2026-05-28 | verified |
| Native context length | developer.nvidia.com ↗ | 2026-05-28 | verified |
| Context on AWS Bedrock | docs.aws.amazon.com ↗ | 2026-05-28 | verified |
| Max output (TheRouter) | docs.aws.amazon.com ↗ | 2026-05-28 | verified |
| Training data | build.nvidia.com ↗ | 2026-05-28 | verified |
| Post-training | developer.nvidia.com ↗ | 2026-05-28 | verified |
| License | www.nvidia.com ↗ | 2026-05-28 | verified |
| Supported languages | aws.amazon.com ↗ | 2026-05-28 | verified |
| PinchBench | developer.nvidia.com ↗ | 2026-05-28 | verified |
| AIME 2025 | aws.amazon.com ↗ | 2026-05-28 | unknown |
| MMLU-Pro | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | 2026-08-25 | verified |
| GPQA (with tools) | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | 2026-08-25 | verified |
| Arena-Hard V2 | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | 2026-08-25 | verified |
| RULER | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | 2026-08-25 | verified |
| TerminalBench 2.0 | NVIDIA NeMo Nemotron 3 Super evaluation recipe ↗ | 2026-08-25 | verified |
| SWE-Bench Verified | aws.amazon.com ↗ | 2026-05-28 | unknown |
| NVIDIA releases Nemotron 3 Super — open hybrid MoE for agentic reasoning | developer.nvidia.com ↗ | 2026-05-28 | verified |
| AWS Bedrock launches Nemotron 3 Super with fully managed inference | aws.amazon.com ↗ | 2026-05-28 | verified |
| DeepInfra adds Nemotron 3 Super with rapid inference | deepinfra.com ↗ | 2026-05-28 | verified |
| What makes LatentMoE different from standard MoE? | developer.nvidia.com ↗ | 2026-05-28 | to verify |
| What context window does Nemotron 3 Super actually support? | docs.aws.amazon.com ↗ | 2026-05-28 | to verify |
| How does Nemotron 3 Super compare to NVIDIA's previous models? | developer.nvidia.com ↗ | 2026-05-28 | to verify |
| Is Nemotron 3 Super suitable for single-turn chatbot use? | llm-stats.com ↗ | 2026-05-28 | to verify |
| What tools/params are supported via TheRouter API? | docs.aws.amazon.com ↗ | 2026-05-28 | to verify |