Ant Group Releases a Large AI Model Anyone Can Download and Use Commercially
Ant Group's inclusionAI team has published the weights of its newest large language model, Ling-3.1-flash, on Hugging Face under an MIT license. MIT is a permissive license: anyone can download the files, change them, and use them inside a commercial product without paying a fee or signing a separate agreement.
The team, which describes itself on Hugging Face as the home of Ant Group's artificial general intelligence work, presents Ling-3.1-flash on its model card as a reasoning model built for complex real-world tasks, aimed at software agents, search, office work and software engineering. The card says it is an upgrade of the earlier Ling-3.0-flash.
The card puts the model at 560 billion parameters in total, with 25 billion of them active for each token of text it processes. The two numbers are not in conflict: this is a mixture-of-experts design, which sends each piece of text through a small part of the network rather than all of it. The stated context window is 1 million tokens, roughly the amount of text the model can keep in view at once.
The card reports results on benchmarks including Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE and HealthBench Professional, and says the model performs well on them. These results are Ant Group's own, and no independent evaluation of Ling-3.1-flash has been published.
The published weights mean the trained model itself has been released, not the data it was trained on. For information about the training data, the model card links to a publicly available summary of its contents.
Running the model is a separate question from licensing it. The card's recommended setup for running it is eight Nvidia H200 accelerators, or four GB300s, which is data-center hardware rather than anything that fits under a desk.
