Ant Group open-sources 560B-parameter Ling-3.1-flash AI model

Ant Group has released the weights for Ling-3.1-flash, opening a model with roughly 560 billion total parameters to developers who want to run it on their own infrastructure. The model activates around 25 billion parameters per token and supports a context window of up to one million tokens, according to the company’s official documentation.¹

The release positions Ling-3.1-flash as a mixture-of-experts model for coding, agent workflows, office productivity and other tasks involving long documents or external tools. Ant’s documentation describes it as part of an open-source Ling family built around long-context modelling, multimodal understanding and agent-style reasoning.¹

A large model with a smaller active footprint

The headline parameter count is not the same as the amount of computation used for every token. The mixture-of-experts design routes work to selected parts of the model, with about 25 billion parameters active at a time. That approach aims to preserve broad capacity without requiring every expert to run for every request.¹

TechNode reports that Ant’s deployment work includes a mixed-precision approach that reduces the model-weight memory requirement from about 534GB to 292GB. In the company’s reported Hopper GPU configuration, that change cuts the deployment requirement from eight GPUs to four.²

The practical benefit is obvious: large open models are only useful to independent developers if the hardware burden is manageable. A four-GPU configuration is still far beyond a typical desktop, but it is a more realistic target for well-equipped labs and teams than an eight-GPU setup.

Long context is the other headline feature

Ant’s official documentation lists a context window of up to 1 million tokens for Ling-3.1-flash, while noting that its official service currently exposes 256,000 tokens. That distinction matters because local deployment and hosted access may offer different limits.¹

The model is aimed at long-horizon software development, everyday office work and agent tasks. Developers can find the open model resources through the InclusionAI and Hugging Face ecosystem, while deployment support includes frameworks such as SGLang and vLLM.¹²

Our opinion

Ling-3.1-flash is interesting less because 560 billion is a dramatic number and more because Ant is pairing that scale with a practical deployment story. The memory reduction will not make the model consumer-friendly overnight, but it shows where open-model competition is heading: bigger capacity, selective activation and increasingly serious attention to the hardware bill. The one-million-token headline also needs the hosted 256K caveat attached, but the underlying push towards longer, locally controlled workflows is significant.