概述
它的故事既展現了對長文本效率的真正架構貢獻,也顯示了中型實驗室在尖端模型上競爭的困難程度。
重點摘要
- AI21 Labs 於 2017 年在特拉維夫由 Yoav Shoham、Ori Goshen 和 Mobileye 創辦人 Amnon Shashua 共同創立。
- 2021 年 8 月發布的 Jurassic-1 Jumbo 擁有 1,780 億個參數,以及包含多詞標記(multi-word tokens)的 25 萬個標記詞彙表。
- 2024 年 3 月發布的 Jamba 在每個區塊中混合了一層注意力層與七層 Mamba 層,並在每個標記上使用其 520 億個參數中的 120 億個。
- 在 256,000 個標記的上下文長度下,Jamba 論文報告其 KV 快取為 4 GB,而 Mixtral 為 32 GB,Llama-2 70B 則為 128 GB。
- AI21 在 2023 年 11 月完成由 Google 和 Nvidia 支持的 2.08 億美元 C 輪融資後,估值達到 14 億美元。
- 2026 年 5 月,AI21 停止了 Jamba 的開發,裁減了大部分員工,並將重心重新聚焦於其 Maestro 代理平台。
深入探討
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.
04工作範例
估算為何 Jamba 的長文本記憶體較小
寫下論文中的原始 Jamba 版面:4 個區塊,每個區塊 8 層,且每個區塊有 1 個注意力層,總計 32 層中有 4 個注意力層。
請注意,只有注意力層會儲存會隨輸入長度增長的 KV 快取;Mamba 層則保留固定大小的狀態。
與同深度的密集 Transformer(全部 32 層皆使用注意力)比較:帶快取的層數從 32 降至 4,減少了八倍。
請查閱論文表 1:Jamba 在 256K token 時需要 4 GB KV 快取,而 Mixtral(32 層、使用 grouped‑query attention)則需要 32 GB。
確認比例:32 GB 除以 4 GB 等於 8,與仍使用注意力的層比例相符。
對 Jamba-1.5-Large 也做同樣計算,其論文報告在 256K token 時為 9 GB,與 Llama-3.1 70B 的 80 GB 進行比較。
它顯示了什麼
學習者會看到 Jamba 的記憶節省幾乎全部來自於將每八層注意力層中的七層換成 Mamba 層,並可直接以已發表的 KV 快取表驗證八倍的數字。
現實世界的實施
Jurassic-1 Jumbo 於 2021 年 8 月透過 AI21 Studio 開放測試版發布,讓開發者可透過 API 使用 1780 億參數模型,並可從 50 到 100 個範例微調自訂版本。
Wordtune 為 AI21 的第一款產品,提供句子改寫的免費增值寫作輔助工具;公司於 2025 年 4 月停止開發,轉而專注企業客戶。
Jamba-1.5-Large 可在單機配備八張 80 GB GPU,使用 2024 年 Jamba-1.5 論文中描述的 ExpertsInt8 量化方法,處理 256,000 token 的上下文。
Jamba Reasoning 3B 於 2025 年 10 月在 Apache 2.0 授權下發布,是一個 30 億參數模型,設計可在筆記型電腦與手機上運行,並支援 250,000 token 的上下文。
Maestro 為 AI21 的代理規劃與編排平台,至 2026 年 5 月成為唯一活躍的產品線,屆時 Nebius 同意將其整合至雲端服務。
AI21 實驗室的未來
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.
付諸實踐
為什麼這很重要
- 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
- 成本與預算. 商業條款和部署選項會影響長期成本和風險。
- 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。
留意
- 發佈公告可能會超過實際生產工作流程的穩定性。
- API 定價或政策轉變可能會在一夜之間打破假設。
- 單一供應商依賴性增加了鎖定和遷移成本。
後續步驟
- 使用您自己的任務和資料集評估提供者。
- 在整合之前查看隱私、安全和法律條款。
- 維護跨模型或供應商的後備計劃。
- 監控發行說明,以便路線圖的變更不會讓團隊感到意外。
資料來源與延伸閱讀
- arXivJamba:混合 Transformer‑Mamba 語言模型
- arXivJamba-1.5:大規模混合 Transformer‑Mamba 模型
- AI21 Labs宣布 AI21 Studio 與 Jurassic-1 語言模型
- VentureBeatAI21 Labs 訓練大型語言模型以匹敵 OpenAI 的 GPT-3
- SiliconANGLE生成式 AI 新創公司 AI21 Labs 獲得額外投資人支持,以 2.08 億美元完成 C 輪融資
- IEEE Spectrum小型語言模型成為新趨勢
- CalcalistNvidia 正就收購 AI21 Labs 進行深入談判
- CalcalistAI21 Labs 解僱多數員工並停止 Jamba 開發
不斷探索
常見問題
AI21實驗室是什麼?
AI21 Labs 是一家成立於 2017 年的以色列 AI 公司,開發了 Jurassic 語言模型、Wordtune 寫作助手以及 Jamba 系列,這是首批將 Transformer 注意力機制與 Mamba 狀態空間層相結合的生產級模型。其發展歷程既展現了對長文本效率的實際架構貢獻,也反映了中型實驗室在尖端模型領域競爭的艱難。
AI21 Labs 的 Jamba 模型因將 Transformer 與其他哪種架構結合而聞名?
Jamba 將 Transformer 注意力層與 Mamba 狀態空間層(加上專家混合層)交錯,以實現高效率的長上下文處理。
將 Mamba 層添加到 Transformer 的主要效率優勢是什麼?
Transformer 注意力成本隨長度呈二次方增長,而 Mamba 狀態空間層則接近線性縮放,從而可以使用更少的記憶體實現很長的上下文。
AI21 Labs 推出了哪些受歡迎的消費者寫作工具?
Wordtune 是 AI21 的人工智慧重寫和音調調整工具,是其最著名的消費產品之一。
AI21 Labs 表示其成立於哪一年?
AI21 的官方「關於我們」頁面指出該公司成立於 2017 年。
AI21 Labs 的哪位共同創辦人也因共同創立 Mobileye 而聞名?
AI21 Labs 共同創辦人 Amnon Shashua 也是自動駕駛公司 Mobileye 的共同創辦人。