Ant Bailing releases Ling-3.1-flash-1019 and its base model
The AMW Read
The release incrementally updates Bailing's model lineup, while permissive base weights and application-driven training have segment-level relevance for agent builders.
Ant Bailing releases Ling-3.1-flash-1019 and its base model
Ant Group's Bailing model team released Ling-3.1-flash-1019 on October 28, alongside Ling-3.1-flash-base. The update targets personal agents that handle complex workflows and multistep tasks. According to company-provided material published by QbitAI, its Artificial Analysis Intelligence Index v4.3.2 score rose three points to 44. AutomationBench-AA increased from 62% to 68%, while AA-Briefcase rose from roughly 1,400 to 1,524 Elo. The base model is available as BF16 weights under an MIT license.
The release connects general-purpose model development with persistent personal-assistant workflows. Lingguang 2.0, upgraded a day earlier, combines cross-session memory, long-running task scheduling and independent cloud hosts. The model team describes application feedback as an input to further training. This links model competition to the ability to advance tasks through tools and changing information. The base model uses a hybrid linear mixture-of-experts architecture, with about 560 billion total parameters and 25 billion activated per token. That design targets model capacity, long-sequence processing and per-step computation together.
For builders, the MIT-licensed base weights provide a starting point for continued pretraining, domain adaptation and post-training around their own tools and workflows. Evaluation should separate the reported benchmark gains from sustained task completion. The supplied material does not report production completion rates or operating costs for the personal-agent workflows. Those gaps matter when assessing assistants expected to keep working after a user leaves a conversation.