Jamba2 3B

A smaller deployment option

AI21 positions this model for local and enterprise tasks where a larger model is unnecessary. The January release pairs it with Jamba2 Mini, which uses a much larger mixture-of-experts network.Introducing Jamba2

Practical guidance. Choose the 3B release when hardware limits matter, then verify extraction accuracy and instruction compliance on the documents it will process. A smaller checkpoint can fit more deployment settings, but its published context limit is not a memory guarantee.

Training beyond the base checkpoint

The release uses extended mid-training followed by supervised instruction tuning, preference optimization and reinforcement learning. The published setup includes mixed reward signals and a progression toward longer sequences.AI21-Jamba2-3B official model card

Download and serve the weights

The weights are published under Apache 2.0. AI21's current model table lists no hosted API endpoint for Jamba 3B, so the jamba-mini API alias must not be used as its identifier.AI21-Jamba2-3B official model cardAI21 Jamba versions, aliases and API deprecations