Skip to main content
AI21 Labs logo

AI21 Labs

Enterprise-grade Jamba models and the Maestro agent orchestration platform

chatbots#llm#enterprise-ai#long-context#rag
Free plan Free trial Claimed API Self-hosted Teams
Toolglade’s take

AI21's Jamba models are a solid enterprise choice for long-context RAG and private, self-hosted deployment, with genuine open weights and broad cloud availability, and Maestro adds a model-agnostic orchestration layer. Buyers should weigh that Jamba is not frontier-class, that AI21 appears to be pivoting away from standalone model sales toward Maestro, and that its funding and acquisition situation carries some uncertainty; confirm current product availability and pricing directly.

About AI21 Labs

AI21 Labs builds the Jamba family of hybrid SSM-Transformer models, offering 256K-token context, efficient inference, and open weights for private deployment, plus Maestro, a model-agnostic orchestration platform for reliable enterprise agents. It targets regulated enterprises needing long-context RAG and grounded, structured outputs. Jamba is not frontier-class, and AI21 appears to be shifting focus toward Maestro, so current product availability should be verified.

AI21 Labs, founded in Tel Aviv in 2017, builds enterprise-focused AI centered on the Jamba model family and the Maestro orchestration platform. Jamba uses a hybrid SSM-Transformer architecture, combining Mamba-style state-space layers with transformer layers, to deliver very long context (256K tokens) and efficient inference, with strengths in retrieval-augmented generation, grounded question answering, function calling, and structured JSON output. The current models, Jamba 1.6 and Jamba 1.7 (released July 2025), are available via AI21 Studio, as open weights on Hugging Face for private on-premises or in-VPC self-hosting, and on cloud marketplaces including Amazon Bedrock. Alongside the models, AI21 Maestro is a model-agnostic planning and orchestration layer that works with OpenAI, Anthropic, and AI21 models and adds requirements, file search, and web search to enforce format, tone, and content constraints. Notably, reporting in 2026 indicates AI21 has been de-emphasizing standalone model sales in favor of Maestro, so buyers should confirm current API availability directly. AI21 targets enterprises, especially regulated industries needing private deployment, long-document workflows, and reliable agents. Its funding picture requires care: roughly $336 million is confirmed through a Series C that reached $208 million at a $1.4 billion valuation (late 2023), while a frequently cited $300 million Series D from Google and NVIDIA and a total near $636 million appear reported but not clearly closed. Reports of NVIDIA acquisition talks in 2026 are rumor-stage. Jamba is a strong long-context, self-hostable option but is not frontier-class, and the company's strategic transition adds some uncertainty.

TL;DR

AI21 Labs builds the Jamba family of hybrid SSM-Transformer models with 256K context and open weights, plus Maestro, a model-agnostic agent orchestration platform. It targets regulated enterprises needing long-context RAG, grounded outputs, and private deployment. Jamba is efficient and self-hostable but not frontier-class, and AI21 appears to be pivoting toward Maestro, adding some product-continuity uncertainty. Confirmed funding is about $336M through Series C.

Company overview

AI21 Labs was founded in 2017 in Tel Aviv by Amnon Shashua, Ori Goshen, and Yoav Shoham. It focuses on enterprise AI, with a history that also includes the consumer writing product Wordtune.

The company has raised roughly $336 million confirmed through a Series C that reached $208 million at a $1.4 billion valuation in late 2023, with investors including Google, NVIDIA, and Intel Capital. A widely cited $300 million Series D and a total near $636 million appear reported but not clearly closed, and 2026 reports of NVIDIA acquisition talks are rumor-stage.

Product features

Jamba 1.6 and 1.7 use a hybrid SSM-Transformer architecture for a 256K-token context window and efficient inference, with strengths in RAG, grounded Q&A, function calling, and structured JSON output. Models are available via AI21 Studio, as open weights on Hugging Face, and on Amazon Bedrock.

AI21 Maestro is a model-agnostic orchestration layer that works across AI21, OpenAI, and Anthropic models, adding requirements, file search, and web search to build reliable agents.

Target market

Enterprises, especially in regulated industries, needing private or self-hosted long-context models, grounded and structured outputs, and reliable agent orchestration.

Buyer personas

End users

Enterprise developers and data teams building RAG, document-processing, and agent workflows.

Buyers

Engineering and platform leaders at enterprises, particularly in regulated sectors requiring private deployment.

Key influencers

ML engineers, security and compliance stakeholders, and enterprise architects.

Ideal customer profile

A regulated or data-sensitive enterprise needing long-context, grounded, self-hostable models and an orchestration layer for reliable agents.

Funding & performance

Roughly $336 million confirmed through a Series C that reached $208 million (originally $155 million in August 2023, extended by $53 million in November 2023) at a $1.4 billion valuation. Investors include Google, NVIDIA, Intel Capital, and others. A reported $300 million Series D and a total near $636 million are widely cited but appear not to have closed; treat as unconfirmed.

Pros & cons

Pros

  • Hybrid SSM-Transformer gives efficient 256K long context
  • Open weights enable private on-prem and in-VPC self-hosting
  • Clear enterprise features: grounding, function calling, structured output
  • Competitive quality-per-cost versus similar-size open models
  • Broad availability across AI21 Studio, Hugging Face, and Bedrock
  • Maestro adds a model-agnostic orchestration layer
  • Backed by major strategic investors including Google and NVIDIA

Cons

  • Strategic pivot toward Maestro creates uncertainty about long-term API availability
  • Jamba is not frontier-class on top-end reasoning
  • Smaller developer mindshare than OpenAI, Anthropic, or Google
  • Primarily text-focused with limited native multimodal
  • Funding and valuation picture is muddied by unconfirmed rounds
  • Enterprise self-hosting requires more engineering effort
  • Shifting pricing and product lineup complicate planning

Pricing plans

Jamba open weights
Free (self-hosted)
  • Download from Hugging Face
  • 256K context
  • Private on-prem or in-VPC deployment
API (AI21 Studio)
From ~$0.22/1M tokens
  • Jamba Mini and Large models
  • Usage-based per-token billing
  • Grounding and structured output
Maestro / Enterprise
Custom
  • Model-agnostic agent orchestration
  • Requirements, file and web search
  • Enterprise support and deployment

Key features

API
Team collaboration
Self-hosted
Multi-language
Integrations
Amazon Bedrock, Hugging Face, AI21 Studio, OpenAI and Anthropic models (via Maestro)
Input types
text
Output types
text
Best For
Long-context RAG, Grounded, structured enterprise outputs, Private and self-hosted deployment, Building reliable AI agents

Compare key features

View all alternatives →
Feature
AI21 Labs
Kimi
Doubao
Pricing
Freemium
Freemium
Freemium
Free plan
Yes
Yes
Yes
Free trial
Yes
No
No
API
Yes
Yes
Yes
Self-hosted
Yes
Yes
No
Team support
Yes
No
No

Frequently asked questions

What is Jamba?+

Jamba is AI21's family of open models using a hybrid SSM-Transformer (Mamba plus Transformer) architecture, offering a 256K-token context window and efficient inference, tuned for enterprise RAG and grounded, structured outputs.

Can I self-host AI21's models?+

Yes. Jamba models are available as open weights on Hugging Face for private on-premises or in-VPC deployment, which is a key draw for regulated enterprises.

What is AI21 Maestro?+

Maestro is a model-agnostic planning and orchestration platform for building reliable enterprise agents; it works with AI21, OpenAI, and Anthropic models and supports requirements, file search, and web search.

How much has AI21 Labs raised?+

Roughly $336 million is confirmed through a Series C that reached $208 million at a $1.4 billion valuation in late 2023. A frequently cited $300 million Series D and a total near $636 million appear reported but not clearly closed; treat those as unconfirmed.

Is the AI21 Studio API still available?+

Reporting in 2026 suggests AI21 has been shifting focus toward Maestro and de-emphasizing standalone model sales. Confirm current API availability directly with AI21 before building on it.

Reviews (0)

Write a review

Pick a rating
Loading reviews…
Compare

Compare AI21 Labs with other AI tools

Side-by-side pages for pricing, features, and best-fit use cases.

All comparisons →

Similar tools you may like