inclusionAI Releases Ling-3.1-flash, a 560B-Parameter MIT-Licensed MoE | Model Release
inclusionAI released Ling 3.1 flash, a native hybrid reasoning model with 560 billion total parameters, 25 billion active parameters per token, official FP8 we…
Published on MyPrivateClaw
Oct 18, 2026, 7:13 AM UTC
Coverage date
Oct 18, 2026
Last updated
Oct 18, 2026, 7:13 AM UTC
News summary
inclusionAI, Ant Group's inclusionAI lab, released Ling 3.1 flash with 560 billion total parameters and 25 billion active parameters per token. The release carries an MIT license and official FP8 weights (F8 E4M3), plus a mixed MXFP4 experts / block FP8 dense deployment recipe. The vendor describes it as among the largest openly licensed MoE releases and the first downloadable Ling 3.1 weights; the exact primary release date was not captured, and the model card serves as the release record. The model is specified with a 1 million token context window and a hybrid linear MoE (BailingMoeV3) architecture: 63 KDA layers plus 9 gated MLA (7:1) layers, 4 dense layers, 512 routed experts plus 1 shared expert, and 8 activated experts per token. Day 0 SGLang support ships in the lmsysorg/sglang:dev ling 3.1 flash image with MTP (NEXTN) speculative decoding and optional Hopper decode optimization…