Ant's Bailing Open-Sources Ling-3.1-flash-1019 for Personal Agent Tasks

Ling-3.1-flash-1019 scored 44 on the Artificial Analysis Intelligence Index v4.3.2, 3 points higher than the previous version. In complex knowledge work, its AA-Briefcase score rose from about 1,400 Elo to 1,524 Elo, and GDPval-AA rose from about 1,618 Elo to 1,644 Elo. AutomationBench-AA increased from 62% to 68%, CritPt from 18% to 26%, and SciCode from 54% to 57%. Bailing says the update improves the model's performance on complex workflows and multi-step tasks.

The release follows the Oct. 27 upgrade of Lingguang 2.0, which Bailing describes as a personal AI assistant that can manage tasks. Lingguang 2.0 uses cross-session memory, long-horizon task scheduling and an independent cloud host to follow up on tasks within the user's authorization and continue managing complex personal affairs even after the user leaves the conversation. That shift from answering questions to handling tasks places new demands on a model: it must understand goals, use tools, process new information and keep a task moving. Bailing cited travel planning and Double 11 shopping as examples, where the assistant must track weather, bookings, budgets, specifications, prices and reminders rather than only produce a single answer.

Ling-3.1-flash-base was designed with those needs in mind. The model uses a hybrid linear MoE architecture with about 560 billion total parameters and about 25 billion activated parameters per token, balancing model capacity, long-sequence processing and per-step compute cost. In 47 benchmarks across eight domains, Ling-3.1-flash-base scored highest on 32 benchmarks and surpassed DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base, according to the article. The base model is available with BF16 base weights under the MIT license, allowing developers to conduct continued pre-training, domain adaptation and post-training using their own knowledge, workflows and tools.

On top of the base model, Ling-3.1-flash and the subsequent 1019 version use post-training to strengthen reasoning, tool use and multi-step execution. Bailing also trained the vision-language model Ling-3.1-flash-VL on Ling-3.1-flash, adding visual understanding to existing reasoning, planning and tool-use capabilities. The VL model has already been used in Lingguang and will be released externally later.

Bailing says Ling-3.1-flash has been aligned with feedback from applications including Lingguang and the Afu Doctor version, improving professional expression and communication. "From real tasks, to real applications," Bailing said, adding that it hopes to improve the model's ability to understand goals, maintain task state and respond to changes through a continuing loop between model development and application practice, while exploring more uses through open source. The company previously released Ling-3.0-flash-Fin, a financial-enhanced model with 124 billion total parameters and 5.1 billion activated parameters, for investment research task flows, and Ling-3.0-flash-Sante, a medical-enhanced model with 124 billion parameters that Bailing says achieved the best performance among models under 1.6 trillion parameters and performance comparable to several flagship models.