Pick AI21 if you want to run open-weight Jamba models privately in your own cloud or build agents on Maestro. Hold off if you need a model vendor with an active roadmap: Calcalist reports Jamba development has stopped and about 70 people remain. Its trust center lists five security certifications.
AI21 Labs is a Tel Aviv AI company founded in 2017 that now sells Maestro, an agent and planning platform, plus lab-as-a-service model training and the open-weight Jamba models with a 256K-token context window. It cut 110 of 180 staff in May 2026 and says Maestro is its focus.
Founded: 2017 · HQ: Tel Aviv, Israel · Team: About 70 (May 2026) · CEO: Ori Goshen · Funding: $200 million (2023 round) · Valuation: $1.4 billion (2023)
About AI21 Labs
AI21 Labs was founded in 2017 in Tel Aviv by Amnon Shashua, Yoav Shoham and Ori Goshen, according to Calcalist's May 2026 report on the company. Its own about page lists Goshen as CEO and Shoham, a Stanford professor emeritus and former Google principal scientist, as chairman. It became a unicorn in 2023 and was for a while Israel's best-known bet on large language models, whose peers in that race included Cohere for enterprise retrieval and, on open weights, Mistral AI.
What AI21 sells has narrowed. Maestro is its agent and planning platform: the docs at docs.ai21.com cover RAG, validated output and user-provided tools over MCP and HTTP, and a December 2025 newsroom post announced Maestro on Amazon VPC. The homepage now leads with "lab-as-a-service" model training, where the company's forward-deployed researchers post-train a specialised model for a customer's task. The Jamba family is still documented as open models (Jamba Large, Jamba2 Mini and Jamba2 3B) with a 256K-token context window and downloadable weights, built on a hybrid of Mamba and Transformer layers. Jamba can run in a customer's own VPC or on-premises, and the docs say the model vendor then has zero visibility of the data.
The retrenchment matters more than the product list. On 19 May 2026 Calcalist reported that AI21 was laying off 110 of its 180 employees, leaving about 70, mostly developers, and that the company would halt development of the Jamba models. It also reported that Wordtune, the consumer writing assistant, had been discontinued in April 2025 (see Wordtune for the tool record and Grammarly as a surviving alternative). On Nebius, AI21 says the outcome was a commercial partnership under which Nebius integrates Maestro into its cloud. Calcalist's sources described an acquihire-style deal for the engineering team, and Nebius told the paper it had agreed to license AI21's technology and that there was no acquihire. Treat the ownership picture as contested until a filing or joint release settles it.
On security, the vendor's trust center lists SOC 2 plus four ISO standards, with Google Cloud Platform and Amazon Web Services named as infrastructure subprocessors. The audit reports themselves sit behind an access request.
Who should look elsewhere: a team that needs a model vendor with a funded roadmap should weigh the frontier labs first (Anthropic, OpenAI), then sovereign-minded options like Aleph Alpha, because AI21's own Jamba line is reported as frozen. Where AI21 still wins is narrow: private deployment of open weights with a long context, and agent orchestration that can be bought through a cloud partner. Its partner and investor trail runs through Together AI, which announced a November 2025 open-model partnership, plus Nvidia and Google (see Nvidia and Google). To shortlist vendors for a specific job, answer a few questions in Smart Match, or browse the companies directory.
Mission
Trustworthy artificial intelligence that powers humanity towards superproductivity.
Products
- AI21 Maestro
- Jamba open models
- Lab-as-a-service training
Compliance
SOC 2 Type II, ISO 27001:2022, ISO/IEC 42001:2023, ISO 27017, ISO 27018
Links
Frequently Asked Questions
How much funding has AI21 Labs raised?
Calcalist reports a $200 million round in 2023 at a $1.4 billion valuation, with no major round since. A reported $300 million raise from Google and Nvidia did not close, and sale talks at a reported $2 billion to $3 billion valuation ended without a buyer for the whole company. The paper names Nvidia, Google, Intel, Samsung Next, 8VC, Coatue and Pitango First among its investors.
What does AI21 Labs sell today?
Maestro, an agent and planning platform, is the main product, alongside lab-as-a-service model training and the open-weight Jamba models. Calcalist reports Jamba development has been halted, and Wordtune ended in April 2025. Existing Jamba weights remain documented and downloadable.
Does AI21 Labs hold SOC 2 and ISO certifications?
Yes to SOC 2 (Type II report on request) and to four ISO standards, one of them the AI-management standard 42001. The trust center also mentions penetration testing and an optional traceless mode. Ask for the reports through the trust center before relying on them in procurement.
What are good alternatives to AI21 Labs?
Choose Cohere for enterprise retrieval and a model line that is still being developed. Choose Mistral AI for open-weight models with an active release cadence. Choose Aleph Alpha when European sovereignty is the requirement, or Together AI to host open models without training your own.
How do I start with AI21 Maestro or Jamba?
Create an API key in the AI21 docs, then use the Python or TypeScript SDK for Maestro. To self-host Jamba, download the weights and deploy them with vLLM in your own cloud. Budget an afternoon for a first working call.
Top Alternatives
- Cohere: Pick Cohere for an enterprise model vendor still developing its models; pick AI21 for privately deployable open-weight Jamba and Maestro agents.
- Mistral AI: Pick Mistral AI if you want open weights that keep shipping; pick AI21 for Maestro orchestration via a cloud partner.
HokAI guides that mention AI21 Labs
- Best AI Writing Assistant in 2026: How to Actually Choose: Grammarly, Wordtune, QuillBot, ProWritingAid and 8 more AI writing assistants compared for 2026, with current prices, free-tier caps and a way to choose.