AI21 Labs

AI21 Labs is an AI lab that offers the Jamba language models, a hybrid Transformer-Mamba architecture, and Maestro, a system for building and deploying AI agents that automate data-intensive business tasks.

Jamba models combine Transformer, Mamba and mixture-of-experts (MoE) layers, according to AI21's original announcement. The documentation's model page lists Jamba Large, Jamba Mini and Jamba 3B, with a maximum context of 256K tokens. AI21 also publishes weights on Hugging Face and says the models can run in a virtual private cloud (VPC) or on premises.

The documentation describes Maestro as a knowledge agent system that uses retrieval over uploaded files, web search and self-validation. Each run builds a tree of calls to language models and other tools, and follows a generate, validate, fix and repeat loop until the defined requirements are met.

Access is through the AI21 Studio API. Maestro is model-agnostic: it can orchestrate AI21's own models or third-party models.

AI21 Labs is for company development and data teams that build knowledge agents over their own documents and, in some cases, need to run the model in a private environment.

Use cases that AI21's documentation and announcements describe:

• Retrieval (RAG) agents over uploaded files, with Maestro, for data-intensive business tasks.
• Answers that must meet explicit requirements. Maestro accepts up to 10 requirements per run, scores each from 0.0 to 1.0 and returns the scores if the call asks for the requirements_result field.
• Private deployment of Jamba models in a VPC, on premises or in a hybrid setup, according to the documentation.
• Local use or small-scale agents, with Jamba2 3B, described as a compact model.

What to have before you start:

• An AI21 Studio account, because the documented access is through the API. Chat calls go to the /studio/v1/chat/completions endpoint, with models such as jamba-large or jamba-mini.
• For retrieval over files, the documents uploaded to the platform.
• To run the weights yourself, attention to each version's license: Jamba2 3B and Jamba2 Mini use Apache 2.0, while earlier versions use the Jamba Open Model License.

AI21 Labs' main features are the Jamba models, with a 256K-token context and private deployment, and Maestro, which validates an agent's answer against requirements before returning it.

• Jamba models: hybrid architecture with Transformer, Mamba and mixture-of-experts layers. The model page lists Jamba Large, Jamba Mini and Jamba 3B, with a 256K-token maximum context.
• Open weights: Jamba2 3B and Jamba2 Mini are on Hugging Face under Apache 2.0. Versions 1.5 and 1.7 use the Jamba Open Model License.
• Deployment: hosted cloud, VPC, on premises or hybrid, according to the documentation.
• Maestro validation loop: generates the answer, scores each requirement from 0.0 to 1.0, fixes whatever is below 1.0, and repeats until the requirements are met or the run reaches its limit.
• Maestro effort control: the budget parameter accepts low, medium or high, and higher values trade latency for more validation. The documentation cites roughly 20, 60 and 100 seconds for complex tasks.
• Maestro tools: retrieval over uploaded files and web search, which could also be used within requirements from September 2025, according to the changelog.
• Model-agnostic Maestro: it orchestrates AI21 or third-party models, such as gpt-4.1 and claude-4-sonnet in the documentation examples.

AI21 Labs is worth it for teams that want knowledge agents with explicit requirements validation and the option to run the models in a private environment.

Pros:

• Maestro returns each requirement's score, from 0.0 to 1.0, when the call asks for requirements_result, which makes validation inspectable.
• Jamba models have a maximum context of 256K tokens, according to the documentation.
• Jamba2 3B and Jamba2 Mini are published under Apache 2.0, and the documentation lists VPC, on-premises and hybrid as deployment options.
• Maestro is model-agnostic and accepts third-party models, so it does not lock the agent into Jamba.

Cons:

• Each Maestro run accepts at most 10 requirements.
• Higher validation levels increase latency: the documentation cites roughly 20, 60 and 100 seconds on complex tasks for low, medium and high.
• The license varies by version: Apache 2.0 for Jamba2, the Jamba Open Model License for versions 1.5 and 1.7.
• The file retrieval API changed: the tool_resources field was marked deprecated in favor of the tools parameter, which means following the changelog.

The main alternatives to AI21 Labs as a model and agent lab are IBM Granite, NVIDIA, Ai2 and Amazon Nova.

• IBM Granite: a family of open language, vision, speech, embedding and guardrail models, with the Apache 2.0 license according to IBM. Granite 4.0 also uses a hybrid architecture with Mamba layers.
• NVIDIA: the open Nemotron model family and NeMo, which bundles microservices for fine-tuning, retrieval and guardrails. NeMo Guardrails is a programmable safety layer for agents.
• Ai2 (Allen Institute for AI): Olmo models published with training data, code and checkpoints, and Molmo vision models.
• Amazon Nova: text, image, video, speech and embedding models accessed through Amazon Bedrock, plus Nova Act for browser task automation.

NVIDIA and IBM list guardrails among their features for AI applications, and Amazon Nova lists the browser automation of Nova Act.