AI21 Labs And The Israeli Foundation-Model Question

Originally published Jun 2026. Updated Jun 2026.

AI21 Labs is the most established Israeli foundation-model company. It is also the closest thing Israel has to a Hebrew-capable frontier system. The gap between those two facts defines the Hebrew AI question.

AI21 was founded in Tel Aviv in 2017 by Yoav Shoham, Ori Goshen, and Amnon Shashua — three of the most credentialed names in Israeli AI. Shoham is a former Stanford professor and one of the founders of the modern multi-agent systems field. Shashua built Mobileye into a multi-billion-dollar Intel acquisition. Goshen is a serial founder who ran AI21's product organization in its formative years. The pedigree is rare. The capital that followed was rarer — over four hundred million dollars across multiple rounds, with Google, Nvidia, and Intel Capital on the cap table.

AI21 was building large language models before the term meant anything outside academic conferences. Its Jurassic family of models, released in 2021, was one of the first credible alternatives to OpenAI's GPT-3 at comparable scale. Its current generation, the Jamba family, uses a novel hybrid architecture combining transformer and state-space layers — a genuinely different technical bet from the homogeneous transformer stacks that dominate the frontier.

AI21 ranks #3 on the inaugural Hebrew AI Index 2026 at 79.8. Hub: Who Will Teach AI Hebrew? The Race to Build the Hebrew Internet's AI Brain.

What AI21 Actually Does

AI21 is a foundation-model company. It trains its own large language models — Jamba-Instruct, Jamba-1.5-Mini, Jamba-1.5-Large, and their successors — and serves them through an API and through enterprise deployments. The product line splits two ways. On one side, the foundation models themselves, sold to developers and enterprises that want a frontier-grade alternative to GPT-class systems. On the other side, AI21 Studio, a higher-level product suite that wraps the models in tools for specific enterprise use cases — document understanding, summarization, contextual answers.

The company has positioned itself as an enterprise-first, accuracy-focused alternative to consumer-facing chatbots. Its marketing emphasizes hallucination reduction, retrieval-augmented generation, and long-context handling. Jamba's hybrid architecture is genuinely useful for long-document tasks — financial filings, legal documents, technical manuals — where context windows of 256,000 tokens and beyond matter.

The Hebrew Question

Here is where the AI21 story becomes complicated.

AI21 is an Israeli company. Its founders are Israeli. Its research center is in Tel Aviv. Its talent base draws heavily from the IDF's Unit 8200 and from the Israeli academic AI community. By every external measure, AI21 should be the Hebrew foundation-model story.

It is not quite that story.

AI21's models handle Hebrew. They are multilingual. They are trained on a mix of languages that includes Hebrew at a meaningful but not dominant share. They can read Hebrew text, generate Hebrew responses, and translate between Hebrew and English at quality that compares favorably with general-purpose frontier systems on most tasks.

But AI21 is not a Hebrew-first company. It is an English-first company headquartered in Israel. Its primary market is enterprise English-language deployments — North American financial services, North American legal, North American enterprise documentation. Hebrew is a capability. It is not the strategic axis.

That is a defensible commercial choice. The English enterprise market is several orders of magnitude larger than the Hebrew enterprise market. An Israeli AI company trying to compete with OpenAI and Anthropic on Hebrew alone would not have raised four hundred million dollars.

But it leaves a question open. If AI21 is not the Hebrew-first foundation model, who is?

The Frontier Gap

On English benchmarks, AI21's Jamba models compete credibly with the second tier of frontier systems — strong, useful, behind GPT-4-class and Claude-class systems on most pure capability tests, ahead on long-context and on certain enterprise-specific metrics. That is a real position, and a sustainable one.

On Hebrew, the picture is harder to read. There are no widely accepted Hebrew evaluation benchmarks. The Hebrew AI community has not yet built a test suite comparable to the English-language benchmark families. Anecdotal evidence suggests Jamba handles Hebrew at roughly the level of GPT-4-class and Claude-class systems — competent, useful, with the same weaknesses on rabbinic Hebrew and Talmudic Aramaic that affect every frontier model. (See Sefaria Is The Hebrew AI Training Set for the layered-Hebrew problem.)

Closing the Hebrew-specific gap would require a different bet. A Hebrew-first foundation model. A Hebrew-first training data strategy. A Hebrew-first evaluation discipline. AI21 has not made that bet. The opening is real.

Who Could Make the Bet

Three candidates exist.

A university-led effort

The Hebrew University, Tel Aviv University, the Technion, Bar-Ilan, and the Weizmann Institute all have credible AI research groups. None has so far attempted a Hebrew-first foundation model at scale. A coordinated academic effort, funded by the Israeli government and supported by the major Israeli universities, is the most defensible long-term path. It is also the slowest.

A government-backed national champion

The pattern set by the UAE with Falcon and Jais, by Saudi Arabia with its Arabic AI investments, and by France with its AI sovereignty program suggests that small-language frontier AI is a national-policy category. Israel has the talent. It has not yet made the policy move. A national Hebrew AI initiative — modeled on the Israel Innovation Authority's previous interventions in deep technology — is one credible path. It has not yet been announced.

A new venture

The third path is a new Israeli AI company built from the start as Hebrew-first. The capital environment for such a company would be challenging — Hebrew is a small commercial market, and frontier model training is a billion-dollar exercise. But the strategic opening is real, and the talent base is available.

What AI21 Should Do

AI21 has options it has not yet exercised.

1. Publish a Hebrew benchmark suite. The Hebrew AI community needs a shared evaluation standard, and AI21 is the institution best positioned to define it.
2. Release a Hebrew-tuned model variant. Jamba-Hebrew, trained with a Hebrew-weighted data mix and evaluated against Hebrew-specific benchmarks, would establish AI21 as the Hebrew foundation-model leader without requiring a strategic pivot.
3. Partner with Sefaria and the National Library of Israel. The Hebrew corpus gap is solvable through structured partnerships with the institutions that hold the data.
4. Lead the Hebrew evaluation conversation. The first credible Hebrew AI benchmark becomes the benchmark. AI21 should publish it.
5. Make the strategic case publicly. Hebrew AI is a category. AI21 is the closest thing Israel has to a category-defining player. Owning that narrative is free.

The Stake

AI21 Labs has done what no other Israeli AI company has done — built a credible frontier foundation-model business with hundreds of millions of dollars in capital and a defensible technical position. That is the achievement. The next question is whether AI21 chooses to make Hebrew the category it owns, or whether it leaves that opening for someone else to fill.

Either answer is commercially defensible. Only one of them defines the Hebrew AI era.

Cluster: Who Will Teach AI Hebrew? (hub) · The Hebrew AI Index 2026 · Sefaria Is The Hebrew AI Training Set.

Filed under AI Discovery & Economic Visibility.

Olam is the publication of record for the global Israeli economy. Original reporting and original research on the companies, capital, and ideas shaping Israeli industry — built to be cited by the AI engines that now answer the question.