Nemotron 3 Nano 30B

​

About the Provider

​

Model Quickstart

• A valid Qubrid API key
• Access to the Qubrid inference API
• Basic knowledge of making API requests in your preferred language

from openai import OpenAI

# Initialize the OpenAI client with Qubrid base URL
client = OpenAI(
    base_url="https://platform.qubrid.com/v1",
    api_key="QUBRID_API_KEY",
)

# Create a streaming chat completion
stream = client.chat.completions.create(
    model="nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
    messages=[
      {
        "role": "user",
        "content": "Explain quantum computing in simple terms"
      }
    ],
    max_tokens=8192,
    temperature=0.3,
    top_p=1,
    stream=True
)

# If stream = False comment this out
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)
print("\n")

# If stream = True comment this out
print(stream.choices[0].message.content)

<think>
The user wants a simple explanation of quantum computing. Let me break this down 
into digestible concepts — classical bits vs qubits, superposition, entanglement, 
and practical applications...
</think>

Quantum computing is a fundamentally different way of processing information.

**Classical vs Quantum:**
- Classical computers use bits that are either 0 or 1
- Quantum computers use qubits that can be 0, 1, or both simultaneously (superposition)

**Key principles:**
- Superposition: A qubit exists in multiple states at once until measured
- Entanglement: Two qubits can be linked so the state of one instantly affects the other
- Interference: Used to amplify correct answers and cancel incorrect ones

**Why it matters:**
Quantum computers can solve certain problems exponentially faster than classical 
computers — particularly in cryptography, drug discovery, and optimization tasks 
that would take classical computers thousands of years.
...

​

Model Overview

• With 31.6B total parameters but only 3.2B active per forward pass, it delivers up to 3.3× higher throughput than comparable models.
• It achieves state-of-the-art accuracy on reasoning, coding, and agentic benchmarks including SWE-Bench, GPQA Diamond, and AIME 2025.
• The model supports up to 262K token context length and features configurable reasoning depth with thinking budget control.

​

Model at a Glance

​

When to use?

• You need agentic AI systems and multi-agent orchestration
• Your application requires complex reasoning and problem-solving tasks
• You are working on code generation, debugging, and optimization
• You need function calling and tool integration
• Your use case involves long-document analysis and RAG applications
• You are solving mathematical reasoning and STEM tasks
• You need enterprise chatbots with deep reasoning capabilities
• Your application requires financial analysis and decision support

​

Inference Parameters

​

Key Features

• Hybrid Mamba-Transformer MoE Architecture: Combines Mamba-2 layers, MoE layers, and GQA attention for optimal efficiency and reasoning performance.
• 3.3× Faster Inference: Delivers higher throughput than Qwen3-30B-A3B with better accuracy, using only 10% of total parameters per forward pass.
• 1M Token Context Window: Supports long-horizon tasks including long-document analysis, RAG pipelines, and multi-turn conversations.
• Configurable Reasoning Depth: Supports reasoning ON/OFF modes with thinking budget control for predictable inference costs.
• Native Tool Calling: Built-in function calling with schema validation and tool-integrated reasoning.
• FP8 Quantization: Reduces memory requirements and enables faster inference on supported hardware.
• Fully Open: Weights, datasets, and training recipes are publicly available.

​

Benchmark Performance

​

Limitations

• VRAM Requirements: Requires 32GB+ VRAM for FP8 and 60GB+ for BF16 deployments.
• Architecture Maturity: Hybrid architecture is less tested in production than pure transformer models.
• MMLU Performance: May underperform on vanilla MMLU compared to harder benchmark variants.
• FlashInfer Dependency: FlashInfer backend requires CUDA toolkit for JIT compilation.
• Community Tooling: New architecture may have limited community tooling support compared to established models.

​

Summary

• Its hybrid Mamba-Transformer MoE architecture delivers 3.3× faster inference than comparable models with only 3.2B active parameters.
• The model supports up to 262K token context, configurable reasoning depth, native tool calling, and FP8 quantization.
• It achieves state-of-the-art results on SWE-Bench, GPQA Diamond, and AIME 2025 benchmarks.
• Weights, datasets, and training recipes are fully open under the NVIDIA Open Model License.