Mamba-Chat is the first chat language model based on a state-space model architecture, not a transformer.

The model is a fine-tune of Albert Gu's and Tri Dao's model Mamba-2.8B from their paper Mamba: Linear-Time Sequence Modeling with Selective State Spaces.

Check our our Github repository for training and inference code.

The prompt format is the zephyr format:

<|user|> {user_message}
<|assistant|> {assistant_message}
<|user|> {user_message}
<|assistant|> {assistant_message}