--- title: "A Beijing nonprofit, not a tech giant, open-sourced a bilingual large model" date: 2026-11-11 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-baai-aquila-open-llm language: en --- # A Beijing nonprofit, not a tech giant, open-sourced a bilingual large model > BAAI's Aquila (悟道·天鹰) line shows how a Chinese research institute built and gave away an open bilingual foundation model — and what that means for developers outside China. Most of the Chinese AI models that made headlines in 2023 arrived wrapped in press releases, not download links. The ones that mattered commercially stayed behind APIs and login walls. So when a government-backed Beijing research institute dropped a bilingual large model (大模型) with open weights, the move drew less attention than it deserved. The lab was not selling a product. It was giving away infrastructure. ## Key takeaways - The Beijing Academy of Artificial Intelligence (BAAI, 北京智源人工智能研究院) released Aquila (悟道·天鹰), a from-scratch open bilingual (Chinese–English) large model (大模型); the second generation, Aquila2, followed on 12 October 2023. - BAAI is a nonprofit founded in November 2018 with backing from the Beijing municipal government and China's Ministry of Science and Technology — it does not sell commercial products. - Unlike the headline 1.75-trillion-parameter WuDao 2.0 (悟道2.0), whose weights were never publicly released, the Aquila family was distributed openly through BAAI's FlagOpen ecosystem and Hugging Face. - The Aquila2-34B model was trained on roughly 1.8 trillion bilingual tokens, according to BAAI's technical report (arXiv, August 2024). - BAAI's broader open stack — FlagAI, FlagEval, and the BGE embedding family — has become widely used plumbing for other developers' models. ### What Aquila actually is Aquila is a series of open large models (大模型) developed at BAAI as a deliberately bilingual base, built for mixed Chinese and English text rather than fine-tuned from a Western model. The first generation appeared in June 2023 as part of the institute's WuDao 3.0 (悟道3.0) announcement, and Aquila2 arrived in October 2023 with 7-billion, 34-billion, and an experimental 70-billion-parameter variant. The Aquila2-34B checkpoint is the best documented. BAAI's technical report describes grouped-query attention, rotary position embeddings, a 100,000-token vocabulary tuned for Chinese–English mixing, and bfloat16 mixed-precision training. The same report states the model was trained on roughly 1.8 trillion bilingual tokens. Weights were published through BAAI's ModelHub and Hugging Face, and the project sits inside the FlagOpen open-source ecosystem rather than a commercial product page. ### Why a nonprofit did this BAAI (北京智源人工智能研究院) is a private non-enterprise legal entity — a "new type of R&D institution" — inaugurated in November 2018 with support from the Beijing Municipal Science and Technology Commission and China's Ministry of Science and Technology. It is not a startup and does not ship paid software. Its stated mission is fundamental AI research and building a shared innovation ecosystem, and its governing model resembles a foundation more than a company. That structure explains the open-weight instinct. Where a commercial lab weighs every release against a revenue plan, BAAI positions its FlagOpen stack as "the Linux of large models" — a common base others build on. FlagAI, its training-and-serving framework, has joined the Linux Foundation, a signal that the institute wants the tooling adopted broadly rather than monetized directly. ### The WuDao contrast To see why Aquila matters, compare it with BAAI's own earlier headline model. WuDao 1.0 launched in March 2021 as China's first large-scale pretrained model system. WuDao 2.0 followed on 1 June 2021 as a 1.75-trillion-parameter mixture-of-experts multimodal model that, at release, was the largest neural network publicly disclosed. It was a demonstration of scale — but its weights were never released to the public. WuDao 3.0, presented in June 2023, marked a strategic pivot: away from one giant showcase model and toward modular components that could be openly released. The language side of that system was branded "WuDao Aquila," and the vision side "WuDao Vision." In other words, BAAI's arc ran from a closed, record-setting spectacle to open, usable building blocks. Aquila is the part of that story a developer can actually download and run. ### What developers actually get The practical value is not Aquila alone but the surrounding stack. FlagOpen bundles FlagAI (model training and serving), FlagEval (a foundation-model evaluation platform covering 800+ open and closed models across 50+ capability dimensions), FlagPerf (AI hardware benchmarking), and FlagEmbedding, the home of the BGE embedding family. Even teams that never deploy Aquila often depend on BAAI's tooling: BGE became Hugging Face's most-downloaded model family of 2023, according to the institute, and BAAI reports its open-source model downloads have exceeded 640 million. For a Chinese or bilingual developer, that combination — an open base model plus evaluation and embedding infrastructure, all under permissive terms — lowers the cost of building a customized product without handing data to a closed provider. ### The honest ceiling Aquila is not a frontier-class challenger to the largest proprietary systems, and BAAI has not marketed it as one. Its strength is as a capable, open, bilingual base that organizations can self-host, audit, and fine-tune. The benchmark conversation around it is quieter than around commercial flagships, which is itself the point: an open model's value is sovereignty, cost control, and inspectability, not leaderboard supremacy. ## Honest limitations - I rely on BAAI's own published figures (its technical report and project pages) and secondary encyclopedic summaries; independent third-party benchmark comparisons of Aquila against current frontier models are sparse, so performance claims here are framed qualitatively rather than as hard rankings. - The exact open-source license terms differ across Aquila checkpoints, and I did not independently audit each one; readers should verify the per-model license on ModelHub or Hugging Face before any commercial use. - BAAI's download and adoption counts are vendor-reported, not independently audited. - This article focuses on the Aquila and FlagOpen line; it does not cover BAAI's Emu multimodal models or the BGE embedding family in depth (the latter is examined separately). ## What readers can do now - If you need a bilingual Chinese–English base model you can self-host and fine-tune, pull Aquila2-34B from BAAI's ModelHub or Hugging Face and benchmark it on your own tasks before committing. - Use FlagEval to compare any candidate model against 800+ open and closed models on 50+ capability dimensions, instead of trusting a single vendor leaderboard. - If your application needs retrieval or search, evaluate BAAI's BGE embedding models as a permissively licensed, widely deployed alternative to proprietary embedding APIs. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-baai-aquila-open-llm Free to quote with attribution and a link to the original.