Service update · Ant Group
Ant Ling open-sources Ling-3.1-flash, a ~560B MoE with 25B active
- Review updated
On October 14, 2026, Ant's Ling team open-sourced Ling-3.1-flash under the MIT license on Hugging Face and ModelScope. It has about 560B total parameters, about 25B active per token and up to 1M context. With MXFP4 experts and block-FP8 dense layers, weights shrink from about 534GB to 292GB, and a tested Hopper setup dropped from 8 GPUs to 4. SGLang and vLLM support it.