Ling 3.1 Flash 1019: open weights, stronger agents, heavier token use

Artificial Analysis measured Ling 3.1 Flash 1019 at 44 on its Intelligence Index v4.3, three points above the 41 it gave the original Ling 3.1 Flash in its earlier test, and its results post shows the update uses considerably more output to get there. The model is an open-weights reasoning model from Ant Ling with 560B total parameters, 25B of them active for each token, and a 1M-token context window. It is released under the MIT license on Hugging Face and is also served through InclusionAI's own API.

The reported gains fall in two areas. On agentic tests, where a model has to carry out multistep work, its Elo rating on AA-Briefcase v1.1 rose 124 points to 1521 and on GDPval-AA v2.1 rose 36 points to 1653, while its AutomationBench-AA score went from 62% to 68%. On scientific reasoning it scored 43% on Humanity's Last Exam, up from 39%, and 26% on CritPt, up from 18%.

The cost of those gains is length. Ling 3.1 Flash 1019 used 145k output tokens per Intelligence Index task against 103k for its predecessor, more than 40% more, and Artificial Analysis says that extra output is the main reason its cost per task, $1.49, is higher than Ling 3.1 Flash's. Its list price is $0.30 per million input tokens and $0.90 per million output tokens, so longer answers show up directly in the bill. Artificial Analysis still puts it below models such as Mistral Large 4 Preview and GLM 5.3 on cost per task.

For anyone choosing between the two versions, the measured trade is a three-point index gain and better agentic results in exchange for substantially more output per task. The post does not give the predecessor's cost per task, so the dollar size of the increase is not stated.