Bedrijven GIDS

AI21 Labs

AI21 Labs is an Israeli AI company founded in 2017 that built the Jurassic language models, the Wordtune writing assistant and the Jamba family, the first production-scale models to mix Transformer attention with Mamba state-space layers.

Luister naar deze gidsAI-verteld · ~8 min

Op deze pagina9 minuten lezen
  1. Overzicht
  2. Belangrijkste inzichten
  3. Diepe duik
  4. Estimate why Jamba's long-context memory is small
  5. Implementatie in de echte wereld
  6. De toekomst van AI21 Labs
  7. Het in de praktijk brengen
  8. Bronnen en verdere literatuur
  9. Blijf verkennen
  10. Veelgestelde vragen

Overzicht

Its story shows both a real architectural contribution to long-context efficiency and how hard it is for a mid-sized lab to compete on frontier models.

Belangrijkste inzichten

  1. AI21 Labs was founded in Tel Aviv in 2017 by Yoav Shoham, Ori Goshen and Mobileye founder Amnon Shashua.
  2. Jurassic-1 Jumbo, released in August 2021, had 178 billion parameters and a 250,000-token vocabulary that included multi-word tokens.
  3. Jamba, published in March 2024, mixes one attention layer with seven Mamba layers per block and uses 12 billion of its 52 billion parameters per token.
  4. At a 256,000-token context, the Jamba paper reports a 4 GB KV cache versus 32 GB for Mixtral and 128 GB for Llama-2 70B.
  5. AI21 reached a $1.4 billion valuation in November 2023 after a $208 million Series C round backed by Google and Nvidia.
  6. In May 2026 AI21 halted Jamba development, cut most of its staff and refocused on its Maestro agent platform.

Diepe duik

AI21 Labs is a Tel Aviv company that builds large language models, the text-predicting neural networks behind chatbots and writing tools. It was founded in 2017 by Stanford computer scientist Yoav Shoham, Ori Goshen and Amnon Shashua, the Hebrew University professor who also founded Mobileye. Shoham and Goshen have served as co-CEOs and Shashua as chairman, according to Calcalist. The company is notable for two things: it was one of the first firms outside the largest US labs to offer a GPT-3-scale model to developers, and it was the first to ship a production model that combines Transformer attention layers with Mamba state-space layers.

The company's first product was Wordtune, a consumer writing assistant that suggested alternative phrasings for a sentence. VentureBeat reported in August 2021 that AI21 then had about 40 employees and $34.5 million in venture funding. That same month AI21 opened AI21 Studio, a developer platform, and released Jurassic-1. The largest version, Jurassic-1 Jumbo, had 178 billion parameters (the adjustable numbers a model learns during training), about 3 billion more than GPT-3. AI21's announcement highlighted a 250,000-token vocabulary, at least five times larger than most vocabularies of the time, which included multi-word tokens such as named entities and common phrases. A token is the unit of text a model reads and writes; packing more meaning into each token means fewer tokens per document, which lowers latency and cost. VentureBeat reported the Jumbo model was trained on 300 billion tokens of English web text, including Wikipedia, news, StackExchange and OpenSubtitles.

Funding followed the generative AI boom. A Series C round closed at $208 million in November 2023, valuing AI21 at $1.4 billion and bringing total funding to $336 million, according to SiliconANGLE. Investors in that round included Google and Nvidia, with Intel Capital and Comcast Ventures joining for the extension. By then the product line included the Jurassic-2 models, task-specific models for jobs such as summarization, Contextual Answers for question answering over a company's own documents, and Wordtune Read for summarizing long texts.

The company's most important technical contribution is Jamba, described in a paper posted to arXiv in March 2024. A standard Transformer uses attention, which compares every token with every earlier token and stores a key-value cache (the KV cache) of past tokens in memory. That cache grows linearly with the length of the input, so very long documents become expensive. Mamba is a selective state-space model: it reads tokens one at a time and carries a fixed-size summary state forward, so its memory does not grow with context length, but on its own it can be weaker at precisely recalling earlier text. Jamba interleaves the two. Each block has eight layers, one attention layer to seven Mamba layers, and every second layer swaps the usual feed-forward network for a mixture-of-experts layer with 16 experts, of which two are used per token. The result is a model with 52 billion total parameters but only 12 billion active for any single token.

The efficiency numbers in the paper are concrete. At a 256,000-token context in 16-bit precision, Jamba needs a 4 GB KV cache, against 32 GB for Mixtral and 128 GB for Llama-2 70B. It fits on a single 80 GB GPU with about 140,000 tokens of text, and on long contexts the authors measured roughly three times the throughput of Mixtral. Quality was competitive rather than leading: Jamba scored 67.4 on the MMLU knowledge benchmark against 70.6 for Mixtral. The first release was a base model with no instruction tuning or safety alignment, published under the Apache 2.0 license, and the authors warned it was not meant for end users without further adaptation.

Jamba-1.5, published in August 2024, scaled the design up. Jamba-1.5-Large has 398 billion total and 94 billion active parameters; Jamba-1.5-Mini has 52 billion total and 12 billion active. Both were released under the Jamba Open Model License. On the RULER long-context test the paper reports an effective context of 256,000 tokens for both, with Large averaging 95.7. The KV cache at 256,000 tokens is 9 GB for Large, compared with 80 GB for Llama-3.1 70B and 252 GB for Llama-3.1 405B. On general benchmarks the models trail the strongest peers in places: Jamba-1.5-Large scores 80.0 on MMLU against 83.6 for Llama-3.1 70B. In October 2025 AI21 added Jamba Reasoning 3B, a small hybrid model that IEEE Spectrum reported runs at more than 17 tokens per second with its full 250,000-token context.

Compared with peers, AI21 occupied a middle position. It never matched the scale or benchmark leadership of OpenAI, Anthropic or Google, and it did not have the distribution of Meta's open Llama models. Its argument was efficiency on long documents, which matters for enterprise retrieval, contract review and summarization. The hybrid idea itself spread: the Jamba paper is widely cited as evidence that a small share of attention layers can restore the recall that pure state-space models lack.

The business side has been difficult. Calcalist reported annual revenue of roughly $50 million that barely changed after 2023, and headcount falling from 260 in 2023 to about 200 at the end of 2025. AI21 shut down Wordtune development in April 2025. In late December 2025 Calcalist reported that Nvidia was in advanced talks to acquire the company for $2 billion to $3 billion, mainly for its staff, but no sale happened. In May 2026 AI21 announced it would cut 110 of its 180 employees, halt development of the Jamba family and concentrate on Maestro, its agent orchestration platform, under a commercial agreement with the cloud provider Nebius. Nebius described the arrangement as a technology license, while Calcalist reported that a large part of the engineering team was expected to move to Nebius.

In practice, using AI21's work today means one of two things. The open Jamba weights remain available and can be served with common inference software; the hardware floor ranges from a laptop for Jamba Reasoning 3B, to one 80 GB GPU for the original Jamba, to an eight-GPU server for Jamba-1.5-Large. Teams choosing Jamba should weigh its long-context memory savings against the fact that its developer has stopped training new versions, so no further updates or fixes should be expected.

04Uitgewerkt voorbeeld

Estimate why Jamba's long-context memory is small

  1. Write down the original Jamba layout from the paper: 4 blocks of 8 layers each, with 1 attention layer per block, so 4 attention layers out of 32.

  2. Note that only attention layers store a KV cache that grows with input length; Mamba layers keep a fixed-size state.

  3. Compare with a dense Transformer of the same depth where all 32 layers use attention: the cache-bearing layers drop from 32 to 4, an eight-fold reduction.

  4. Check the paper's Table 1: Jamba needs 4 GB of KV cache at 256K tokens, while Mixtral, a 32-layer model with grouped-query attention, needs 32 GB.

  5. Confirm the ratio: 32 GB divided by 4 GB equals 8, matching the share of layers that still use attention.

  6. Repeat for Jamba-1.5-Large, whose paper reports 9 GB at 256K tokens, and compare it with 80 GB for Llama-3.1 70B.

Wat het laat zien

The learner sees that Jamba's memory saving comes almost entirely from replacing seven of every eight attention layers with Mamba layers, and can verify the eight-fold figure directly against the published KV cache table.

Implementatie in de echte wereld

Jurassic-1 Jumbo, released through the AI21 Studio open beta in August 2021, gave developers API access to a 178-billion-parameter model and let them train custom versions from as few as 50 to 100 examples.

Wordtune, AI21's first product, suggested rephrasings of sentences as a freemium writing aid; the company stopped developing it in April 2025 to focus on enterprise customers.

Jamba-1.5-Large can process a 256,000-token context on a single machine with eight 80 GB GPUs by using the ExpertsInt8 quantization method described in the 2024 Jamba-1.5 paper.

Jamba Reasoning 3B, released in October 2025 under Apache 2.0, is a 3-billion-parameter model designed to run on laptops and phones while handling a 250,000-token context.

Maestro, AI21's agent planning and orchestration platform, became its only active product line by May 2026, when Nebius agreed to integrate it into its cloud services.

De toekomst van AI21 Labs

The open question AI21 helped pose is how much attention a language model really needs. Jamba showed in 2024 that one attention layer in eight could keep long-context recall while cutting cache memory by roughly an order of magnitude, and its successor paper reported a confirmed 256,000-token effective context on RULER. Whether hybrids keep that advantage as context windows grow toward millions of tokens, and how they compare with improved attention-only designs that compress the KV cache, remains an active research question rather than a settled one. For the company itself, the direction changed in 2026. Calcalist reported in May 2026 that AI21 would stop developing Jamba models, reduce staff to about 70, and focus on Maestro, its system for planning and running multi-step AI agent workflows, with Nebius as a major partner. That makes AI21 a case study in a broader pattern: smaller labs moving from training foundation models to building orchestration and application layers on top of models from others. The Jamba weights and papers remain public, so the architecture can continue to be studied and extended by other groups even without new releases from AI21.

Het in de praktijk brengen

Waarom het ertoe doet

  • Leveranciersstrategie. Roadmaps van leveranciers beïnvloeden welke functies uw team vervolgens kan bouwen.
  • Kosten en budget. Commerciële voorwaarden en implementatieopties zijn van invloed op de kosten en risico's op de lange termijn.
  • Risico en veiligheid. Bedrijfsprikkels bepalen productgebreken, veiligheidshouding en openheid.

Pas op voor

  • Lanceringsaankondigingen kunnen de stabiliteit in echte productieworkflows overtreffen.
  • API-prijzen of beleidswijzigingen kunnen van de ene op de andere dag de aannames doorbreken.
  • De afhankelijkheid van één leverancier verhoogt de lock-in- en migratiekosten.

Volgende stappen

  1. Evalueer providers met behulp van uw eigen taken en datasets.
  2. Controleer de privacy-, beveiligings- en juridische voorwaarden vóór de integratie.
  3. Onderhoud een noodplan voor alle modellen of leveranciers.
  4. Houd de release-opmerkingen in de gaten, zodat wijzigingen in de routekaart teams niet verrassen.

Bronnen en verdere literatuur

  1. arXivJamba: A Hybrid Transformer-Mamba Language Model
  2. arXivJamba-1.5: Hybrid Transformer-Mamba Models at Scale
  3. AI21 LabsAnnouncing AI21 Studio and Jurassic-1 Language Models
  4. VentureBeatAI21 Labs trains a massive language model to rival OpenAI's GPT-3
  5. SiliconANGLEGenerative AI startup AI21 Labs finds additional backers to close Series C funding round at $208M
  6. IEEE SpectrumSmall Language Models Are the New Rage
  7. CalcalistNvidia in advanced talks to acquire AI21 Labs
  8. CalcalistAI21 Labs lays off most employees and halts Jamba development

Blijf verkennen

Gratis nieuwsbrief

Ontvang de dagelijkse AI-briefing

Drie verifieerde AI-verhalen elke weekdagmorgen, geschreven in eenvoudig Engels. Gratis voor altijd, geen advertenties.

Eén e-mail elke weekdag. Afmelden met één klik. We verkopen of delen je adres nooit.

Test jezelf

Take the AI21 Labs quiz

Directe feedback op elke antwoord, en een deelbare certificaat met een verifieerbare ID zodra je een cursus hebt afgerond.

Quiz starten

Ondersteun gratis AI-onderwijs. AI Understanding is een 501(c)(3) non-profit — geen advertenties, geen betaalmuur, nooit. Doe een donatie

Veelgestelde vragen

What is AI21 Labs?

AI21 Labs is an Israeli AI company founded in 2017 that built the Jurassic language models, the Wordtune writing assistant and the Jamba family, the first production-scale models to mix Transformer attention with Mamba state-space layers. Its story shows both a real architectural contribution to long-context efficiency and how hard it is for a mid-sized lab to compete on frontier models.

Het Jamba-model van AI21 Labs valt op door het combineren van de Transformer met welke andere architectuur?

Jamba verweeft Transformer-aandachtslagen met Mamba-statusruimtelagen (plus een mix van experts) voor efficiënte afhandeling van lange contexten.

Wat is het belangrijkste efficiëntievoordeel van het toevoegen van Mamba-lagen aan een Transformer?

De aandachtskosten voor transformatoren groeien kwadratisch met de lengte, terwijl Mamba-toestandsruimtelagen dichter bij lineair schalen, waardoor zeer lange contexten met minder geheugen mogelijk zijn.

Welke populaire schrijftool voor consumenten is gemaakt door AI21 Labs?

Wordtune, AI21's AI-aangedreven hulpmiddel voor herschrijven en toonaanpassing, is een van de bekendste consumentenproducten.

In what year does AI21 Labs say it was founded?

AI21’s official About page states the company was founded in 2017.

Welke medeoprichter van AI21 Labs staat ook bekend als medeoprichter van Mobileye?

Amnon Shashua, medeoprichter van AI21 Labs, was ook medeoprichter van het autonoom rijdende bedrijf Mobileye.