NVIDIA Nemotron 3 Super (120B MoE)

Nemotron 3 Super is NVIDIA's open-source 120B-total / 12B-active Mixture-of-Experts Hybrid Mamba-Transformer model, released March 11, 2026. Designed specifically for complex agentic AI systems — autonomous coding, cybersecurity triaging, and long-form multi-step research. Delivers 5× higher throughput vs dense models of comparable quality.

Why Run Nemotron 3 Super on Clore.ai?

Nemotron 3 Super's MoE architecture means only 12B parameters are active per forward pass — so you get frontier-level reasoning at the compute cost of a mid-sized model. On Clore.ai you can rent a single RTX 5090 (32GB) or a pair of RTX 4090s and run it with full INT4/FP4 quantization at production speeds.

Key numbers:

• 120B total parameters, 12B active (Latent MoE)

• Hybrid Mamba-Transformer architecture (first in Nemotron line with MTP Layers)

• 1M token context window

• Pre-trained in NVFP4 — native NVIDIA FP4 quantization

• 5× throughput vs comparable dense models

• NVIDIA Nemotron Open Model License — open weights with commercial use

Hardware Requirements

Config	VRAM	Clore.ai Cost	Notes
FP4 (native)	1× RTX 5090 32GB	~$3.50–5/hr	Fastest; native NVFP4
INT4	2× RTX 4090 24GB	~$4–6/hr	Strong option
INT4	1× A100 80GB	~$20/hr	Full INT4, single GPU
INT8	4× RTX 4090	~$8–12/hr	Near-full quality
BF16 full	4× H100 80GB	~$24–40/hr	Training / full fidelity

Best value on Clore.ai: 2× RTX 5090 (available from ~$7/hr) for BF16 full-precision inference.

Quick Start: vLLM + Nemotron 3 Super

# Pull the vLLM Docker image (NVFP4 support requires vLLM >= 0.7.3)
docker run --gpus all --rm -it \
  -p 8000:8000 \
  -v /root/.cache:/root/.cache \
  vllm/vllm-openai:v0.7.3 \
  --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 \
  --quantization fp4 \
  --max-model-len 32768 \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.92

For multi-GPU (2× RTX 4090 in INT4):

docker run --gpus all --rm -it \
  -p 8000:8000 \
  -v /root/.cache:/root/.cache \
  vllm/vllm-openai:v0.7.3 \
  --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 \
  --quantization awq_marlin \
  --max-model-len 65536 \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.90

SGLang (Alternative — Faster MoE Serving)

For production-grade MoE throughput, SGLang's RadixAttention gives 2–5× better throughput vs vLLM on MoE models:

docker run --gpus all --rm -it \
  -p 30000:30000 \
  -v /root/.cache:/root/.cache \
  lmsysorg/sglang:latest \
  python -m sglang.launch_server \
    --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 \
    --tp 2 \
    --quantization fp8 \
    --context-length 131072 \
    --port 30000

Deploy on Clore.ai: Step-by-Step

1. Rent a GPU

Go to clore.ai/marketplace:

• Filter: RTX 5090 or RTX 4090 × 2+

• Sort by price (spot orders are 20–40% cheaper)

• Minimum: 32GB VRAM total (FP4); 48GB for INT8; 80GB for BF16

2. Launch Container

In the Clore.ai dashboard, select Custom Docker and enter:

Image: vllm/vllm-openai:v0.7.3
Ports: 8000
Command: --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 --quantization fp4 --max-model-len 32768

Or use the one-liner SSH launch:

ssh root@<clore-server-ip> "docker run --gpus all -d \
  -p 8000:8000 \
  -v /root/.cache:/root/.cache \
  --name nemotron3 \
  vllm/vllm-openai:v0.7.3 \
  --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 \
  --quantization fp4 \
  --max-model-len 32768 && echo 'Started'"

3. Test the API

curl http://<server-ip>:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Write a Python function to scrape GitHub issues and categorize them by severity."}
    ],
    "max_tokens": 2048,
    "temperature": 0.1
  }'

Agentic Use Case: Multi-Agent Coding Pipeline

Nemotron 3 Super is purpose-built for multi-agent workflows. Here's a minimal example using the OpenAI-compatible API:

from openai import OpenAI

client = OpenAI(
    base_url="http://<server-ip>:8000/v1",
    api_key="none"
)

def planning_agent(task: str) -> str:
    """High-level task decomposition."""
    response = client.chat.completions.create(
        model="nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
        messages=[
            {"role": "system", "content": "You are a senior engineering lead. Break down complex tasks into concrete sub-tasks with acceptance criteria."},
            {"role": "user", "content": f"Decompose this task: {task}"}
        ],
        max_tokens=1024,
        temperature=0.0
    )
    return response.choices[0].message.content

def coding_agent(subtask: str) -> str:
    """Code implementation."""
    response = client.chat.completions.create(
        model="nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
        messages=[
            {"role": "system", "content": "You are an expert Python engineer. Write production-quality code with tests."},
            {"role": "user", "content": subtask}
        ],
        max_tokens=2048,
        temperature=0.1
    )
    return response.choices[0].message.content

# Example: autonomous feature implementation
plan = planning_agent("Build a REST API for user authentication with JWT")
print("Plan:", plan)
code = coding_agent(f"Implement step 1 from this plan: {plan}")
print("Code:", code)

Benchmarks (March 2026)

Benchmark	Nemotron 3 Super	DeepSeek V3	Llama 4 Maverick
HumanEval	92.1%	90.8%	88.4%
MATH-500	89.3%	90.2%	84.7%
SWE-bench Verified	65.2%	61.4%	55.8%
MMLU	88.7%	87.2%	86.1%
Throughput (tok/s)	1,840	410	890

Throughput measured on 2× H100 80GB with INT4 quantization.

Monitoring & Production Tips

# Watch GPU memory and utilization
watch -n2 nvidia-smi

# Check vLLM throughput stats
curl http://localhost:8000/metrics 2>/dev/null | grep vllm

# Docker logs (live)
docker logs -f nemotron3

# If OOM: reduce max_model_len or increase tensor-parallel-size

Recommended settings for production on Clore.ai:

• --max-model-len 32768 for most workloads (saves VRAM, covers 95% of requests)

• --gpu-memory-utilization 0.90 (leave 10% buffer for MoE routing overhead)

• --enable-chunked-prefill for better latency on long inputs

• Enable spot orders for 30–40% cost savings on batch workloads

Cost Comparison

Provider	Config	$/hr
Clore.ai (spot)	2× RTX 5090	~$5.60
Clore.ai (on-demand)	2× RTX 5090	~$7.00
Azure AI	Hosted API	~$15–20
NVIDIA API	Hosted API	~$12–18

Self-hosting on Clore.ai is 2–3× cheaper than managed API for sustained workloads.

Related Guides

• vLLM Serving — production LLM server with OpenAI-compatible API

• SGLang — faster MoE throughput with RadixAttention

• DeepSeek V4 — upcoming 1T-parameter open model

• CrewAI — build multi-agent pipelines with role-based agents

• OpenHands — autonomous software engineering agents

• GPU Comparison — pick the right GPU for your workload

---

Last updated: March 16, 2026 | Model released: March 11, 2026 | License: NVIDIA Nemotron Open Model License