Aquila

Overview

Aquila is a family of open-source large language models developed by the Beijing Academy of Artificial Intelligence (BAAI). It was built with an explicit focus on bilingual Chinese-English capability, domestic Data Compliance, and permissive commercial licensing. Rather than being a single hosted chatbot, Aquila is distributed as a set of model weights and training recipes that developers download, run, and fine-tune on their own infrastructure.

Because the project is released in the open, its exACT lineup and capabilities may change over time as new checkpoints and training methods are published. Treat the details below as a General picture rather than a fixed specification.

What it does

Aquila provIDEs the raw building blocks for language APPlications. The base models deliver general text understanding and generation and can be adapted to downstream tasks through continued pre-training or supervised fine-tuning. Dedicated chat checkpoints are tuned for conversational interaction, and code-oriented variants target text-to-code generation, producing snippets in multiple programming languages, including Python.

A notable design goal is native Chinese knowledge. The training mix reportedly gives Chinese corpora a substantial share, drawn from curated internet data and authoritative Chinese Literature sources, so the model does not rely only on translated English knowledge.

Key capabilities

• Bilingual modeling: handles both Chinese and English Prompts and ouTPUts within one model.
• Multiple variants: parameter scales in the billions and tens of billions, plus chat and code-tuned veRSIons.
• Commercial-friendly terms: the code follows Apache 2.0 while the weights are governed by a separate Aquila model license, so users should read the terms carefully.
• Fine-tuning support: training and configuration scripts are included for pre-training, supervised tuning, and reInforcement-learning-style alignment.
• Efficient training tooling: the project references parallel training utilities and optimizations aimed at NVIDIA A100-class hardware.

Who it is for

Aquila suits machine learning engineers and research teams who need a self-hosted bilingual model they can inspect, modify, and deploy inside their own environment. It is also relevant for organizations with data-residency or compliance requirements that rule out sending text to third-party APIs, and for academic groups studying Chinese-centric language modeling or Code Generation.

How to get started

1. Review the model card and license terms to confirm commercial use is perMITted for your case.
2. Obtain the weights and example training or inference scripts from the official project repository.
3. Prepare GPU infrastructure; larger variants need substantial memory and may require multi-GPU setups.
4. Run Inference on Sample prompts, then fine-tune on your own dataset if needed.
5. Evaluate outputs for accuracy, safety, and language quality before production use.

Pricing and access caveat

Aquila itself is distributed at no charge, but running it is not free in practice because compute, stoRAGe, and engineering time are required. Licensing for weights differs from the Apache 2.0 code license, so consult the official agreement before commercial deployment. Hosted availability and version numbers vary by release, so verify current details at the source.