Ant Group publishes Ling-3.1-flash-1019, a 574 GB open model

On 27 October, the Ling team at Ant Group put the open weights of Ling-3.1-flash-1019 on Hugging Face with an MIT licence. Open weights are model files that anyone can download and operate. The files total 573.8 GB. Five days before, Ant published only the score of 44 on the Artificial Analysis index, up from 41. The index owner says that it measured the score itself.

What is in the files

This is a full model, not a small patch. Only 25 billion of its parameters work on each token. A token is a unit of text that a model reads or writes. A version for more training is also available.

The numbers

The score of 44 is measured by others. The increases on single tests come from the model card of Ant. They are not from a test by others. A task costs $1.49 in the test, against $1.01 for the earlier Ling-3.1-flash. The model is last of 48 for verbosity, which is how much text it writes.

What is not known

The model card says that the context is 1M tokens. The config files in the repository set 256K. This is the own analysis of the files by the source, not a statement of Ant. The hosted API can use a different adjustment. Test the limit on your own system.

Insight: Open weights win trust when buyers can check the claim

Open-weight models processed more than 7 in 10 tokens on the Vercel AI Gateway in September. Mistral plans to release the open weights of Large 4 on 31 October. Ant gave its files five days after the score. This reads as a contest for buyer trust: files that anyone can test, with the price paid in more tokens. Buyers get more control and can compare the cost for each task, not the score.