Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

AI21 released Jamba with a 256K context window

AI21’s original Jamba is a downloadable base language model that combines transformer attention with Mamba state-space layers. It supports inputs with long passages of text. [1][3]

Model access
Jamba-v0.1 base weights are downloadable. AI21 later announced a separate Jamba-Instruct model. [1][2]
License
Apache License 2.0 for original Jamba-v0.1 weights. [1]

What the model does

The company released Jamba-v0.1 on March 28, 2024. The base model card describes a 256K-token context and about 12 billion parameters active out of 52 billion total. [1][3]

Why it matters

Jamba offers a way to assess a different architecture on long documents. A large advertised context lets teams submit more material together, while the hardware needed still depends on the workload. [1][3]

Who can use it

Model researchers and infrastructure teams can download the original base checkpoint. Developers evaluating a chat application should distinguish it from the later instruction-tuned Jamba release. [1][2]

Key dates

  1. Jamba-v0.1 base weights released [3]
  2. Separate Jamba-Instruct model announced [2]

Ways to use it

Evaluate the base model on a long-text task and measure whether it can recover details spaced far apart. Check memory use at realistic document lengths alongside answer quality. [1][3]

Model limitations

The full published context and the amount practical on one device can differ. The March checkpoint is a base text model, with later Jamba releases carrying their own model cards and terms. [1][2]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source AI21 Labs model card and license for Jamba-v0.1
  2. Primary source AI21 Labs announcement of Jamba-Instruct
  3. Primary source Introducing Jamba: AI21's Groundbreaking SSM-Transformer Model | AI21
  4. Independent source AI21 Labs Unveils Jamba: The First Production-Grade Mamba-Based AI Model

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked