Introducing Jamba2

Jamba2 is a family of open-source language models aimed at enterprise reliability and controllability, offering a compact, memory-efficient setup that's easy to drop into a production agent stack. It's built on an SSM-Transformer architecture and is designed to handle precise QA and document-based workflows without the heavy thinking token burden of reasoning models.

The model family comes in two versions.

• 3B dense
• Mini (MoE): 12B active, 52B total parameters

Both are released under Apache 2.0 and support a 256K context window. They're available on AI21 Studio and Hugging Face, and AI21 emphasized that the 3B model can run locally on devices like iPhone, Android, Mac, and PC.

On performance, the company reported leading results on IFBench, IFEval, Collie, and the grounding benchmark FACTS. Notably, it maintains high throughput even as context extends to 100K tokens, positioning it with a reliability-to-throughput ratio suited to enterprise QA workflows.

In human evaluations using real enterprise task prompts, Jamba2 Mini outperformed Ministral3 14B in overall output quality and win rate. The evaluation was conducted as a blind pairwise comparison based on factuality, style, constraint adherence, instruction-following, and helpfulness.

Training consisted of mid-training on 500B tokens, followed by a state passing phase, cold-start SFT, DPO, and multi-stage on-policy RL that started with short contexts and expanded to long contexts. The key was combining verifiable rewards with model-based rewards to refine the model into one that follows instructions more accurately and doesn't miss detailed requirements.