Skip to content
Neural Daily

ModelsNews

Ling-3.1-flash-1019: Ant's MIT-licensed 560B model

🇭🇺 Olvasd el magyarul →

Ant Group's inclusionAI team released Ling-3.1-flash-1019 weights: a 560B MoE with 25B active parameters, 1M token context and an MIT license. It is clearly stronger than its predecessor on agentic tests.

By: Mikola Máté Patrik · · 2 min read

Ling-3.1-flash architecture diagram with KDA and Gated MLA blocks, MoE layers and multi-token prediction
Image: inclusionAI · Press kit · Official architecture diagram from the inclusionAI model card · Source

What was released?

inclusionAI, Ant Group's AGI team, published the Ling-3.1-flash-1019 weights on Hugging Face (Hugging Face). It is a further post-trained version of Ling-3.1-flash: same architecture, more post-training. The Ling-3.1-flash-base model was also released for your own fine-tuning and domain adaptation.

  • 560B total, 25B active parameters per token, 1M token context.
  • Hybrid linear MoE: 63 KDA linear-attention and 9 Gated MLA layers in a 7:1 ratio, 512 routed experts with 8 active.
  • FP8 checkpoint, about 574 GB, MIT license.

How much better is it?

  • AA-Briefcase: from about 1400 to 1524 Elo.
  • AutomationBench-AA: from 62% to 68%.
  • CritPt: from 18% to 26%, SciCode: from 54% to 57%.
  • Artificial Analysis Intelligence Index v4.3.2: from 41 to 44.
Bar chart of Artificial Analysis Intelligence Index v4.3.2 scores with Ling-3.1-flash-1019 highlighted at 44
inclusionAI · Press kit · Official chart from the inclusionAI model card · Source

Source: inclusionAI. Artificial Analysis Intelligence Index v4.3.2 standings in the company's chart, with Ling-3.1-flash-1019 at 44.

Context matters here: Artificial Analysis moved to Intelligence Index v5 on Tuesday, where released open weights models trail much further behind. The 44 is on the old v4.3.2 scale.

What to keep in mind

It needs serious hardware: the company recommends 8 H200s or 4 GB300s under SGLang. The model card's SGLang example points to the old Ling-3.1-flash path, so set the model path to inclusionAI/Ling-3.1-flash-1019 by hand, or you will benchmark the previous version. The custom model code requires trust_remote_code.

Why it matters

This is one of the strongest MIT-licensed open models for agentic work, and the base model is freely available too. If you build self-hosted agents and have the infrastructure, it is worth comparing against DeepSeek V4.1 Flash.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles