Ant Group launches AI model Ling-3.1-flash with 560 billion parameters

The research unit InclusionAI, which belongs to Ant Group, has released a new language model. Ling-3.1-flash has a total of 560 billion parameters, of which approximately 25 billion are actually activated for each request. This architecture is designed to save computing power without making major compromises in the quality of responses.

What the model is supposed to do

Ant Group is positioning Ling-3.1-flash for use in so-called agent systems, that is, AI programs that independently complete multi-stage tasks. According to the company, these also include search functions, office applications, and solutions for specialized fields. The model is designed for a context window of up to one million tokens, so it can theoretically process very long texts or documents at once.

In the current two-week testing phase, however, access is limited to a significantly smaller context window. This means that users cannot yet fully exploit the model's capacity. Ant Group provides no information on when the full version will be freely available.

The technology behind the model

The principle of activated parameters, also called Mixture-of-Experts, works like a large specialist library, from which only the relevant books are retrieved for each request, rather than searching through the entire library.

With Ling-3.1-flash, this corresponds to just under 4.5 percent of the total capacity that is actually used per request. This reduces energy and computing costs compared to models that use their full parameter count for each request.

By comparison: models like OpenAI's GPT-4 or Google's Gemini also rely on similar architectures to make large models economically viable to operate. Ant Group's strategy thus fits into an industry-wide trend in which Chinese providers like DeepSeek, Alibaba, and Moonshot AI have pursued comparable approaches in recent years.

Market Positioning

Ant Group is best known for its financial subsidiary Alipay, which is used by more than one billion people in China. With InclusionAI, the corporation expanded its business into foundational artificial intelligence models, a sector that has recently been heavily contested by companies like Alibaba, Baidu, and Tencent.

The release of Ling-3.1-flash as an open-weight model, in which the trained parameters are made publicly accessible, follows a pattern that several Chinese AI providers have recently pursued. This allows developers worldwide to download the model, adapt it, and integrate it into their own applications without relying on a paid programming interface.

For European and Austrian companies that use AI models for customer service, document processing, or software development, the selection of available base models is thereby expanded in addition to existing options from OpenAI, Google, Meta, and Mistral AI. Open models from China like those from DeepSeek or Alibaba are already being tested by individual European companies, since they are often usable without licensing fees and can be operated locally, which can be relevant for data protection concerns. Business associations such as the Austrian Chamber of Commerce are monitoring developments in open-source AI models as part of their digitalization initiatives for domestic enterprises.

Source: TechNode | Original Article