Ant Ling open-sources Ling-3.1-flash-1019: further post-trained 560B/25B-active MoE lifts AA Intelligence Index from 41 to 44, base checkpoint also released

Background and pain point

Open-weight competition has shifted from topping single leaderboards to reliably finishing real multi-step agent tasks. Ant Ling released Ling-3.1-flash in early October (covered here on Oct 3), a 560B-total, 25B-active MoE pitched on cost efficiency. The 1019 version keeps the same base but adds further post-training for reasoning and real-world agent workflows, and for the first time also open-sources the base checkpoint so companies can keep training on their own domain data.

Architecture

According to the Hugging Face model card, Ling-3.1-flash-1019 is a hybrid-linear MoE: 63 KDA linear-attention layers plus 9 gated MLA layers in a 7:1 ratio, using linear attention to cut long-context compute and memory while a few full-attention layers keep retrieval accuracy. It has 512 routed experts and 1 shared expert, activates 8 experts per token, and supports a 1M-token context. Weights are MIT-licensed and usable commercially.

Benchmarks

All figures are vendor-reported: 44 on the Artificial Analysis Intelligence Index v4.3.2, up from 41; AA-Briefcase from about 1400 to 1524 Elo; AutomationBench-AA from 62% to 68%; CritPt from 18% to 26%; SciCode from 54% to 57%. Ant published these in the model card and its official X thread; we did not check each one against Artificial Analysis's own pages this round, and no head-to-head comparison with closed frontier models has been published.

Getting started

Weights are on Hugging Face and ModelScope under the inclusionAI organization. Ant recommends SGLang with the day-0 image lmsysorg/sglang:dev-ling-3.1-flash: TP8 plus EP8 on 8 H200s or TP4 plus EP4 on 4 GB300s, with MTP speculative decoding. On vLLM, enable the ling3 tool-call and reasoning parsers. Recommended sampling is temperature 1.0, top_p 0.95 and top_k 20 with enable_thinking on. At 560B it still needs a multi-GPU server, so individual developers are better off with the online demo or a third-party API.