| Property | Value |
|---|---|
| Model Size | 8B parameters |
| Training Tokens | 3.5T |
| Context Length | 4K |
| License | Apache 2.0 |
| Paper | Link to Paper |
What is mamba2-hybrid-8b-3t-4k?
The mamba2-hybrid-8b-3t-4k is an innovative language model that combines the strengths of Mamba-2 architecture with traditional attention and MLP layers. Developed by NVIDIA using the Megatron-LM framework, this 8B-parameter model represents a hybrid approach to language modeling, trained on an impressive 3.5T tokens with a 4K sequence length.
Implementation Details
This model implements a hybrid architecture that leverages the selective state space model (SSM) from Mamba-2 alongside conventional transformer components. It's built using the Megatron-LM toolkit and supports extensions for longer context lengths up to 32K and 128K.
- Hybrid architecture combining Mamba-2, attention, and MLP layers
- 8B parameters optimized for efficient processing
- 4K sequence length with available extensions
- Built on NVIDIA's Megatron-LM framework
Core Capabilities
- High-quality text generation
- Efficient processing of long sequences
- Balanced performance between attention and state space mechanisms
- Scalable architecture supporting context length extensions
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its hybrid architecture that combines the innovative Mamba-2 selective state space model with traditional attention mechanisms, offering a balanced approach to language modeling while maintaining efficient processing capabilities.
Q: What are the recommended use cases?
The model is particularly well-suited for text generation tasks requiring both long-range dependencies and efficient processing. It's ideal for applications needing robust language understanding while benefiting from the advantages of both SSM and attention mechanisms.