Ant Group

What it is

Ant Group is the organisation whose InclusionAI team released the Ling family of large language models, called Bailing (百灵) in Chinese, in February 2025 ^[1][2]. The family is a set of language models built on a mixture-of-experts (MoE) architecture and released with open weights: they are published on Hugging Face and ModelScope under the MIT licence, which permits reuse and modification ^[3]. Two models were released together and differ mainly in scale — Ling-Lite, a compact model, and Ling-Plus, a much larger one ^[1].

Where it came from

The release was accompanied by a technical report whose title states the project's aim directly: the team scaled a 300-billion-parameter model without premium GPUs ^[1]. The Ling-Plus model card was created on 28 February 2025 ^[2], and the technical report was first submitted to arXiv on 7 March 2025 ^[1]. Taken together, the material describes a project organised around a hardware constraint rather than around the largest available cluster.

Main features

The two models differ sharply in size. Ling-Lite has 16.8 billion total parameters, of which 2.75 billion are activated for any given piece of computation; Ling-Plus has 290 billion total parameters with 28.8 billion activated ^[1]. Both are mixture-of-experts models, meaning only a subset of the parameters is used at a time, which keeps the computational cost of a single forward pass well below what the total parameter count would suggest ^[1].

The report says the team trained a 300-billion-parameter mixture-of-experts model on lower-performance devices while matching comparable dense and mixture-of-experts models ^[1]. To get there, the report contributes methods for architecture and training-process optimisation, for handling anomalies during training, and for making evaluation more efficient ^[1]. It also describes the data side: training data generated from knowledge graphs gave the models stronger tool-use ability than peers, according to the report ^[1]. Tool use here refers to a model calling external functions or programs rather than answering only from its own weights.

Where it sits

Open weights under the MIT licence put the family in the part of the field where models can be downloaded, inspected and adapted rather than only queried through an API ^[3]. The weights are hosted on both Hugging Face and ModelScope ^[3]. The family did not stay fixed after launch: Ling-lite was upgraded to Ling-lite-1.5 on 10 May 2025, and the report and repository track that revision ^[1][3].

For Ant Group, the Ling family is the company's visible entry into foundation-model research, and its distinguishing claim is procedural rather than merely numerical — that a model of roughly 300 billion parameters can be built on less powerful hardware and still be compared with dense and mixture-of-experts models of similar standing ^[1]. That claim, together with the released weights and the published account of training methods and failure handling, is what the project offers to readers and to other teams working under similar compute limits.