Introducing Jamba2

Jamba2: Open Source LLMs for Enterprise Reliability

Skip to Main Menu

Skip to Main Content

Skip to Footer

Back to Blog

-->

Back to Blog

Today, we are introducing Jamba2, an open source family of language models built for maximum reliability and steerability in the enterprise.

Built on our novel SSM-Transformer architecture , and a category leader across grounding and instruction following benchmarks, Jamba2 offers a compact, memory-efficient addition to any production agent stack, able to power precise question answering workflows that don’t call for the heavy “thinking token” overhead of reasoning models.

Jamba2 is available in two model sizes: 3B and Mini (MoE; 12B active, 52B total parameters). We are proud to release both under the Apache 2.0 License , as part of our continued commitment to democratizing access to quality models. We are particularly excited this release includes a 3B model, enabling developers everywhere to download and run this technology right on their own devices, including iPhones, Androids, Macs, and PCs.

Jamba2 model family fast facts

License: Apache 2.0

Model sizes: 3B (dense) 52B/12A

Context window length: 256K

Availability: AI21 Studio , Hugging Face

Maximum reliability and steerability

With over seven years of experience building AI systems for enterprise customers, we know that reliability and steerability are non-negotiables when it comes to choosing a model. 54% of enterprises still cite accuracy as the number one AI-related risk they are working to mitigate—with 30% of surveyed organizations having experienced the negative drawbacks of AI inaccuracies at least once ( McKinsey, 2025 ). Even as models get increasingly sophisticated, there is an enormous, high-stakes gap to be filled when it comes to enterprise-grade accuracy.

We also know, from our work with customers, that not every enterprise workflow requires the high cost and high latency of reasoning models. Especially as more organizations adopt agents and model routers, a strong enterprise AI system should contain a range of model types, with different models called as needed based on task type and budget considerations.

With this in mind, we set out to build a model that delivers precise output with a lean memory footprint. To make “precision” measurable, we focused on two signals that map directly to real deployments: instruction-following benchmarks, which capture how steerable a model is, and grounding benchmarks, which test whether outputs stay faithful to the provided context. Together, they serve as practical indicators of how consistently a model will behave in enterprise knowledge workflows.

Today, we’re bringing that vision to market with Jamba2: compact models that lead on these measures of enterprise reliability. Jamba2 is built to produce grounded answers across a range of source types, including technical manuals, research papers, company policies, and internal knowledge bases, so teams can deploy it as a dependable component in production stacks.

Leader on instruction following and grounding

Jamba2 models excel across the instruction following benchmarks IFBench, IFEval, and Collie, as well as the grounding benchmark FACTS.

Winning reliability-to-throughput ratio

For the enterprise, models that shine on quality, yet choke in production-scale settings, are unusable.

Jamba2 leverages its memory-efficient architecture to maintain high throughput alongside high enterprise reliability, even as context scales to 100K tokens, mirroring the real-world usage we can expect from enterprise QA workflows.

Winning performance on enterprise tasks

In a human evaluated comparison between Jamba2 Mini and Ministral3 14B on a test set of real-world enterprise task prompts, Jamba2 Mini showed a statistically significant advantage in overall output quality and win rate. The tasks included a mix of enterprise QA tasks, instruction-heavy developer prompts, and other common business tasks such as summarization and drafting. This win rate encapsulates human preference, with special attention to factuality, style, constraint-adherence, instruction-following, and helpfulness.

Note on human evaluation methodology : The models were evaluated by a team of content evaluation experts using a side-by-side, blind comparison protocol; left–right ordering was counterbalanced across items to mitigate order and precedence effects. Evaluators reviewed both outputs for both absolute metrics (general quality measures, as well as fine-grained error analysis), as well as human preference.

How we built it

To train Jamba2, we utilized a contemporary LLM post-training pipeline. Following the pre-training phase, we mid-trained Jamba2 on 500B carefully curated tokens, with a higher representation of math and code in the mix, along with high-quality web data and long documents. We completed the mid-training phase with a short state passing phase for the model’s Mamba layers, a method recently introduced for effectively generalizing an SSM’s context length.

Given our midtrained model, we performed cold start SFT to teach the model basic instruction-following and reasoning. We then topped off the model with DPO to further improve its performance as a starting point for on-policy RL. Finally, we ran multiple on-policy RL phases on our model, starting with short-context verifiable rewards and gradually moving to longer context training with a mix of verifiable and model-based rewards.

Carefully aggregating verifiable and model-based rewards was one of the key advancements allowing Jamba2 models to excel on tasks that require following specific instructions and details while adhering to the general intention of the user. As part of this process, we optimized our training infra for efficient on-policy RL training, which combines model inference in both the generation and reward phases of the training.

Getting started with Jamba2

Available for download on Hugging Face or directly on AI21 Studio , researchers and AI enthusiasts alike can enjoy experimenting with Jamba2 and pushing it to support new use cases—we can’t wait to see what you build!

Discover more

Jun 25, 2026

Token spend isn’t going down. You need more than naive routing to manage it

Labs in Front

-->

Jun 24, 2026

Tipping the scales: Merging weak agents into a state-of-the-art deep researcher

Labs in Front

-->

Jun 4, 2026

First scale, then enrich: How the right execution strategy helped us reach...