AI21 Labs iconAI21 Labs

Israeli AI lab behind the Jamba hybrid Mamba-Transformer models, now refocused on its Maestro agent platform after 2026 layoffs

$0.20 Jamba Mini per 1M Input
256K Context (Jamba Reasoning 3B)
2017 Founded

Overview

AI21 Labs is a Tel Aviv AI company founded in 2017 by Yoav Shoham, Ori Goshen and Amnon Shashua. It builds the Jamba models, which mix Mamba state-space layers with Transformer attention to handle long documents cheaply, and Maestro, a platform for building and running knowledge agents (live since July 2025). Jamba is sold as an API (Jamba Mini and Jamba Large), and several versions have open weights on Hugging Face, including Jamba2 Mini (52B), Jamba Large 1.7 (399B) and the small Jamba Reasoning 3B, released in October 2025 under Apache 2.0 with a 256K context. The company changed course in 2026: in May it cut about 110 of roughly 180 staff and, per Globes, stopped selling standalone models to focus on Maestro. Acquisition talks with Nvidia (late 2025) and later with Nebius were reported, and AI21 denied that a deal was near; no acquisition had closed in the sources we found. The pricing page still lists Jamba API prices, and the homepage shows new posts in October 2026, so the company is active, but buyers of the Jamba API should confirm its future with sales.

The Verdict

Who Should Use AI21 Labs?

Best For

  • Enterprises building grounded knowledge agents on Maestro
  • Long-document work where Jamba's hybrid design keeps cost down
  • Teams that want to self-host open Jamba weights

Not Ideal For

  • Teams that need a model API with a long, certain roadmap
  • Frontier-level reasoning or code generation
  • Buyers who want public, self-serve Maestro pricing

What's Great

  • Low API prices: Jamba Mini $0.20 in, $0.40 out per 1M tokens
  • Hybrid Mamba-Transformer design built for long context
  • Open weights for several Jamba models, Apache 2.0 for Jamba Reasoning 3B
  • $10 free trial credit, no credit card

Watch Out For

  • About 60% of staff laid off in May 2026; focus moved from models to Maestro
  • Ownership uncertain after reported Nvidia and Nebius talks
  • Very few public user reviews (no G2 page with reviews found)
  • Maestro pricing is not public

Pricing

View all features & details

Models

  • Jamba2 Mini (52B) and Jamba2 3B, open weights (2026)
  • Jamba Reasoning 3B: Apache 2.0, 256K context (Oct 2025)
  • Jamba 1.7 Mini (52B) and Large (399B), self-hosted deployment (Jul 2025)
  • Jamba 1.5 on Amazon Bedrock and Google Vertex AI

Products

  • Maestro: knowledge agents and agent optimization
  • AI21 Studio and foundation model APIs with SDK
  • Post-training services and Intelligent Gateway

Company

  • Founded November 2017, Tel Aviv
  • $155M Series C at a $1.4B valuation (Aug 2023), with Google and Nvidia
  • May 2026: about 110 of 180 staff cut

Real-World Usage

Named Customers

  • Gong
  • Deel (Emet by Deel testimonial)
  • Wix plans to run its own operations on Maestro

Status in 2026

  • Standalone model sales being wound down in favor of Maestro (Globes)
  • Jamba prices still listed on the official pricing page in November 2026

How It Compares

Feature AI21 Jamba Mistral Small 4 Reka Flash 3
Architecture Hybrid Mamba-Transformer Mixture of experts (119B, 6.5B active) Multimodal (21B open version)
Input price per 1M $0.20 (Jamba Mini) $0.15 $0.10
Output price per 1M $0.40 (Jamba Mini) $0.60 $0.20
Open weights Yes (several Jamba versions) Yes, Apache 2.0 Yes, Apache 2.0 (Flash 3.1)
Company focus in 2026 Maestro agent platform Models and platform Video and vision products

Sources: AI21 Labs · Pricing · Changelog · Hugging Face · Globes (May 2026) · Calcalist (2026) · Wikipedia

Is this your tool? Add a badge