Ling (Ling-Lite and Ling-Plus)

What it is

Ling is a family of large language models built on the Mixture-of-Experts (MoE) architecture, released by InclusionAI, the AI research group at Ant Group. Its Chinese name is Bailing ^[1]. The family launched with two members: Ling-Lite, with 16.8 billion total parameters of which 2.75 billion are activated for any given piece of text, and Ling-Plus, with 290 billion total parameters and 28.8 billion activated ^[1].

MoE is a design in which a model contains many separate sub-networks, usually called experts, and a routing mechanism sends each token to only a few of them. This separates two numbers that are often conflated: total parameters describe how much the model can store, while activated parameters describe how much computation each token actually costs. A model can therefore be very large on paper and still be relatively cheap to run.

Where it came from

The team published a technical report for the family, first submitted on 7 March 2025 ^[1]. The Ling-plus model card on Hugging Face was created earlier, on 28 February 2025 ^[3]. The report's title frames the work around a constraint rather than a capability: it describes scaling a 300B Mixture-of-Experts model without premium GPUs ^[1].

How it was built

The central engineering claim is that the team trained a 300B MoE on lower-performance devices while matching comparable dense and MoE models ^[1]. Getting a model of that size to train on less powerful hardware is largely a problem of efficiency, and the report contributes methods in several areas: optimizing the architecture and the training process, handling anomalies that arise during training, and making evaluation more efficient ^[1]. Training anomalies — loss spikes, unstable gradients and similar failures — become more likely and more expensive as models grow, so procedures for detecting and recovering from them are part of what the report documents.

The team also reports a data-side result: training data generated from knowledge graphs gave the models stronger tool-use ability than peers ^[1]. Tool use means the model can call external functions, search or APIs rather than only producing text, and knowledge graphs — structured stores of entities and their relations — are a way of generating training material that is factual and relational rather than purely conversational.

Release and later versions

The weights are released under the MIT license and distributed on Hugging Face and ModelScope ^[2]. An MIT license is a permissive open-source license, which means the models can be used and modified commercially with few restrictions — a notable choice for a model of this size from a major fintech-affiliated group.

The family did not stay fixed. Ling-lite was upgraded to Ling-lite-1.5 on 10 May 2025, with better reasoning than its predecessor ^[2].

Where it sits

Ling is best understood as an argument about cost as much as a model release. Two things distinguish it: a MoE design that keeps activated parameters far below total parameters at both sizes, and an explicit claim that competitive training is possible on hardware that is not top of the line. The comparisons with dense and MoE peers are the team's own reported results, as are the tool-use findings. What can be checked independently is the release itself — the weights, the license and the report — which is what makes the work useful to others working under similar hardware limits.