Jamba v0.1

Jamba v0.1 is the original base model published in March 2024. Its mixture-of-experts network combines attention and state-space layers, using 12 billion active parameters from 52 billion total.

Why attention and Mamba were combined

AI21 used attention to address recall limitations in a pure state-space model, while retaining Mamba's efficient sequence processing. The release established the Joint Attention and Mamba architecture used by later Jamba models.AI21 introduces the Jamba architecture

This checkpoint is not an instruction assistant

The model card warns that this is a pretrained base release without instruction tuning or moderation. It is intended as a starting point for further training.Jamba-v0.1 official model card

Practical guidance. Do not interpret fluent text completion as reliable chat behavior. Applications that need conversation, tool calls or controlled answers need an appropriate adapted model.

Mamba blocks need separate treatment

The card's 8-bit loading example leaves Mamba blocks outside the quantized modules to avoid quality loss. This makes the recommended loading path different from simply quantizing every linear layer.Jamba-v0.1 official model card

Specifications & access

Jamba v0.1 specifications
DetailValue and conditions
Launch date
AccessDownloadable
Model identifiersJambaai21labs/Jamba-v0.1
Context window262,144 tokens

Maximum position count in the published weight configuration; deployed limits depend on the serving setup.

Jamba-v0.1 configuration
Parameters52 billion total; 12 billion activeJamba-v0.1 official model card
Weight licenseApache 2.0Jamba-v0.1 official model card

Related models

  • Jamba Tiny Dev

    AI21 Labs · Jamba research

    319-million-parameter trained development checkpoint for work on the Jamba architecture.

    Launch date Research releaseDownloadable
  • Original instruction-tuned Mini model with tool calls, JSON output and long text input.

    Launch date Deprecated API; check access detailsDownloadable · API
  • Historical hosted instruction version of the original Jamba architecture.

    Launch date Access not confirmed