Ling 3.0

InclusionAI is Ant Group's research arm and the Ling family, internally 百灵 or Bailing, is its open-weight output. Learn the structure first because nothing on the Hugging Face org page explains it: Ling is the non-thinking line, Ring is the reasoning line, Ming is the multimodal series and LLaDA is a diffusion language model track. All of it ships MIT, which puts Ant alone among Chinese labs shipping at trillion-parameter scale on genuinely unencumbered terms.

Ling-3.0-flash, open-sourced 5 August 2026, is the release most teams should evaluate. 124B total parameters with 5.1B activated non-embedding, a 256K native context and a hybrid attention stack alternating Kimi Delta Attention with Gated MLA layers at a 5:1 ratio over a 1/64-sparse MoE of 512 routed experts. Ling-3.0-tiny followed a day later at 7.9B total and 1.3B active with a 262K context, aimed at tool-using agents. At the top of the range, Ring-2.6-1T carries 1 trillion total parameters with 50B active and adaptive effort levels that trade latency for deeper chains of thought, though serving it means four nodes at tensor-parallel 8 with pipeline-parallel 4 under SGLang.

Two habits set Ant apart from most labs publishing weights. It releases base checkpoints at several training stages, including 30T-token pretrain and midtrain snapshots, which matters if you intend to do your own post-training rather than fine-tune someone's instruct model. And it ships its own quantisations, with BF16, FP8, INT4 and MXFP4 builds all coming from the lab and targeting Hopper and Blackwell directly. The weak point is everything around the models. Compared with Qwen or DeepSeek you will find fewer community fine-tunes, fewer GGUF conversions and less mature third-party tooling, and the English-language reporting is unreliable enough that third-party write-ups of Ling-3.0-flash have circulated a 51B active-parameter figure when the model card says 5.1B.