Chat and reason with NVIDIA Nemotron 3.5 Super, a hybrid Mamba-Transformer mixture-of-experts model with optional step-by-step thinking.

nvidia/nemotron-3-5-super/text

Chat and reason with NVIDIA Nemotron 3.5 Super, a hybrid Mamba-Transformer mixture-of-experts model with optional step-by-step thinking.

Inference

Commercial use

Schema

Playground API Examples

Input

Prompt*

Additional Settings

Customize your input with more control.

Result

Idle

What would you like to do next?

nvidia/nemotron-3-5-super/text

Schema

Input

Prompt*

Additional Settings

Customize your input with more control.

Result

Idle

What would you like to do next?

Input

Prompt*

Additional Settings

Customize your input with more control.

Result

Idle

Hybrid Mamba-Transformer models combine the architectural strengths of both to overcome individual weaknesses:

**1. Complementary Capabilities**
*   **Transformers:** Use self-attention for global context and precise token-to-token relationships, but suffer from quadratic computational complexity as sequence length increases.
*   **Mamba (State Space Models):** Use selective scanning for linear scaling and efficient long-sequence processing, but may lack the fine-grained local precision of attention.

**2. Efficiency and Scalability**
By integrating Mamba layers, hybrids reduce memory usage and inference latency for long contexts while maintaining the high-performance modeling capabilities of Transformers.

**3. Architectural Integration**
Models typically alternate Mamba and Transformer blocks or use them in parallel paths, allowing the network to dynamically leverage global attention for complex dependencies and SSMs for broad sequence traversal.

**4. Goal**
The objective is to achieve state-of-the-art accuracy with significantly improved throughput and lower hardware requirements compared to pure Transformer models.

What would you like to do next?

{
  "output": "Hybrid Mamba-Transformer models combine the architectural strengths of both to overcome individual weaknesses:\n\n**1. Complementary Capabilities**\n*   **Transformers:** Use self-attention for global context and precise token-to-token relationships, but suffer from quadratic computational complexity as sequence length increases.\n*   **Mamba (State Space Models):** Use selective scanning for linear scaling and efficient long-sequence processing, but may lack the fine-grained local precision of attention.\n\n**2. Efficiency and Scalability**\nBy integrating Mamba layers, hybrids reduce memory usage and inference latency for long contexts while maintaining the high-performance modeling capabilities of Transformers.\n\n**3. Architectural Integration**\nModels typically alternate Mamba and Transformer blocks or use them in parallel paths, allowing the network to dynamically leverage global attention for complex dependencies and SSMs for broad sequence traversal.\n\n**4. Goal**\nThe objective is to achieve state-of-the-art accuracy with significantly improved throughput and lower hardware requirements compared to pure Transformer models.",
  "reasoning": null,
  "finish_reason": "stop",
  "usage": {
    "input_tokens": 36,
    "output_tokens": 202
  }
}

Your request will cost $0.01 per 1,000 tokens (input plus output), with a minimum of $0.01 per request.

Input

Prompt*

Additional Settings

Customize your input with more control.

Result

Idle

What would you like to do next?

NVIDIA Nemotron 3.5 Super Text (Large Language Models) API on fal

Input

Prompt*

Additional Settings

Customize your input with more control.

Result

Idle

Hybrid Mamba-Transformer models combine the architectural strengths of both to overcome individual weaknesses:

**1. Complementary Capabilities**
*   **Transformers:** Use self-attention for global context and precise token-to-token relationships, but suffer from quadratic computational complexity as sequence length increases.
*   **Mamba (State Space Models):** Use selective scanning for linear scaling and efficient long-sequence processing, but may lack the fine-grained local precision of attention.

**2. Efficiency and Scalability**
By integrating Mamba layers, hybrids reduce memory usage and inference latency for long contexts while maintaining the high-performance modeling capabilities of Transformers.

**3. Architectural Integration**
Models typically alternate Mamba and Transformer blocks or use them in parallel paths, allowing the network to dynamically leverage global attention for complex dependencies and SSMs for broad sequence traversal.

**4. Goal**
The objective is to achieve state-of-the-art accuracy with significantly improved throughput and lower hardware requirements compared to pure Transformer models.

What would you like to do next?

{
  "output": "Hybrid Mamba-Transformer models combine the architectural strengths of both to overcome individual weaknesses:\n\n**1. Complementary Capabilities**\n*   **Transformers:** Use self-attention for global context and precise token-to-token relationships, but suffer from quadratic computational complexity as sequence length increases.\n*   **Mamba (State Space Models):** Use selective scanning for linear scaling and efficient long-sequence processing, but may lack the fine-grained local precision of attention.\n\n**2. Efficiency and Scalability**\nBy integrating Mamba layers, hybrids reduce memory usage and inference latency for long contexts while maintaining the high-performance modeling capabilities of Transformers.\n\n**3. Architectural Integration**\nModels typically alternate Mamba and Transformer blocks or use them in parallel paths, allowing the network to dynamically leverage global attention for complex dependencies and SSMs for broad sequence traversal.\n\n**4. Goal**\nThe objective is to achieve state-of-the-art accuracy with significantly improved throughput and lower hardware requirements compared to pure Transformer models.",
  "reasoning": null,
  "finish_reason": "stop",
  "usage": {
    "input_tokens": 36,
    "output_tokens": 202
  }
}

Your request will cost $0.01 per 1,000 tokens (input plus output), with a minimum of $0.01 per request.