AI21 Labs

AI21 Labs is an Israeli artificial intelligence company founded in 2017 by Yoav Shoham, Ori Goshen and Amnon Shashua. Headquartered in Tel Aviv, AI21 is one of the longest-established independent foundation model developers and is notable for combining transformer-based language models with structured reasoning components aimed at enterprise reliability. The company's flagship model family is Jamba, a hybrid Mamba-Transformer architecture, and its enterprise agentic platform is Maestro.

History

AI21 emerged from research at Stanford and the Hebrew University of Jerusalem, where co-founder Shoham had previously worked on knowledge representation and game theory. The company's first consumer product, Wordtune, an AI writing assistant, launched in 2020. The Jurassic-1 language model was announced in August 2021, briefly making AI21 one of the few organisations operating models comparable in scale to GPT-3.

Subsequent releases included Jurassic-2 in March 2023, which improved efficiency and added instruction tuning. In March 2024 AI21 introduced Jamba, the first production-grade language model to combine the Mamba state-space architecture with transformer attention layers in a mixture-of-experts configuration. The company has since concentrated development on the Jamba family while keeping Jurassic accessible through its API.

Jamba architecture

Jamba's hybrid architecture interleaves transformer attention blocks with Mamba state-space model blocks, augmented by a mixture-of-experts (MoE) routing layer. The design retains the strong in-context learning of transformers while exploiting the linear-time, memory-efficient scaling of state-space models for long sequences. Jamba models support a 256,000-token context window and are reported by AI21 to deliver inference throughput roughly 2.5 times higher than comparable transformer-only models at similar quality.

The Jamba 1.5 family, released in 2024, includes Jamba 1.5 Mini and Jamba 1.5 Large. Jamba 1.6 was released in March 2025 for private enterprise deployment with reported gains over other open models on enterprise benchmarks, and Jamba 1.7 followed on Hugging Face in August 2025. Open weights for several Jamba variants are distributed under the Jamba Open Model Licence.

Maestro and enterprise AI

Launched in 2025, Maestro is AI21's orchestration system for building knowledge agents in regulated enterprise environments. Maestro combines Jamba models with retrieval, planning, tool-use and policy enforcement layers, and is positioned as an alternative to general-purpose agent frameworks for organisations subject to strict accuracy, auditability and data-residency requirements. AI21 markets Maestro to financial services, healthcare and government customers.

Funding and partnerships

AI21 has raised approximately USD 636 million across seven rounds. A USD 208 million Series C in August 2023 was led by Walden Catalyst, Pitango and SCB10X, with participation from Google and NVIDIA. A USD 300 million Series D in May 2025 was led by NVIDIA and Alphabet. Jurassic and Jamba models are available through AWS Amazon Bedrock, Microsoft Azure AI Foundry, Google Cloud Vertex AI, Snowflake Cortex and Databricks Mosaic AI.

Significance

AI21 occupies a distinctive niche in the foundation model landscape: it operates at frontier scale, ships open weights for selected models, focuses sharply on enterprise reliability and was an early industrial backer of non-transformer architectures. Its work on Jamba contributed to broader 2024–2025 interest in hybrid state-space and attention designs from competitors including Mistral, Together AI and Cartesia.

References

1. AI21 Labs. (2024). Jamba: A Hybrid Transformer-Mamba Language Model. arXiv:2403.19887.
2. AI21 Labs. (2025). Jamba 1.6 Release Notes. AI21 Studio.
3. VentureBeat. (2024). AI21 Labs juices up gen AI transformers with Jamba.
4. Contrary Research. (2025). AI21 Labs Business Breakdown.
5. AWS. (2025). AI21 Labs Models in Amazon Bedrock.