Ant Bailian open-sources its Ling-3.1-flash AI model

Ant Bailian has open-sourced its Ling-3.1-flash AI model. It has about 560B total parameters, with about 25B active per token, and supports up to 1M context. It scored 41 on Artificial Analysis tests, with output at about 210 tokens per second.

For local use, the team used MXFP4 quantization, cutting the model size from about 534GB down to 292GB. On Hopper GPUs, setup dropped from 8 cards to 4, making it easier for teams to run.

For long documents, speed tweaks cut waiting time. In tests, the first token took 604 milliseconds on 8 GPUs, and 4 GPUs handled about 1.36 requests per second.

The model is free to download on Hugging Face and ModelScope. There is also a limited-time 75% off deal, with OpenRouter pricing at $0.075 for input and $0.225 for output.