Ling-3.1 Flash 1019 Goes Open Source

The benchmark gain is notable, but the base checkpoint is the more consequential part of this release: it gives researchers a permissively licensed foundation for domain tuning and further post-training.

• –Post-training lifted AA-Briefcase from roughly 1,400 to 1,524 Elo, AutomationBench-AA from 62% to 68%, CritPt from 18% to 26%, and SciCode from 54% to 57%. [Model card](https://huggingface.co/inclusionAI/Ling-3.1-flash-1019)
• –The 560B-total, 25B-active hybrid-linear MoE combines sparse routing with a claimed 1M-token context window, targeting reasoning and long-horizon agent workflows.
• –Ling-3.1-flash-base supports continued pretraining, supervised fine-tuning, preference optimization, distillation, and MoE research under MIT licensing. [Base model card](https://huggingface.co/inclusionAI/Ling-3.1-flash-base)
• –This is still serious infrastructure, not a laptop model: the released weights are roughly 572B parameters, and the recommended serving setup starts at eight H200 GPUs. [Model card](https://huggingface.co/inclusionAI/Ling-3.1-flash-1019)