Ant Group Unveils Ling-3.0-Flash, an Efficient AI Model for Agents

Chinese fintech giant Ant Group has released Ling-3.0-Flash, a new AI model the company says matches or beats systems two to three times its size on reasoning, instruction-following, and long-context tasks — while activating only a fraction of its parameters per query.

A leaner mixture-of-experts design

Ling-3.0-Flash comes from inclusionAI, Ant Group’s AI research lab, and totals 124 billion parameters. Like other mixture-of-experts models, it activates only a small subset of those parameters for any given token — in this case just 5.1 billion, according to Ant Group’s July 27 announcement. That is a steeper compression than the company’s earlier Ling-2.6 lineup, and Ant Group says the new model still matches or surpasses its own roughly 1-trillion-parameter Ling-2.6 flagship on most of the benchmarks it tested — using about one-twelfth the active parameters.

The model relies on a hybrid attention architecture that alternates two techniques — Kimi Delta Attention, adapted from Moonshot AI’s research, and multi-head latent attention — at a 5-to-1 ratio, which Ant Group says makes long documents cheaper to process. Its context window natively covers 256,000 tokens, with the company stating a roadmap toward 1 million.

Built for AI agents, not chat

Ant Group is positioning Ling-3.0-Flash less as a chatbot and more as fast infrastructure for AI agent pipelines — what it calls a “high-speed execution node” meant to plan, call tools, and self-correct across multi-step tasks. The company says it trained the model across more than 10,000 interactive environments and added a caching system that cuts time-to-first-response by 60 to 80 percent on long inputs.

Free for now, open weights later

Ling-3.0-Flash is currently free to use through OpenRouter and Vercel AI Gateway, through August 3, after which standard usage pricing applies. Ant Group says the model’s weights will follow as an open release once that window closes, echoing Chinese labs like Moonshot AI, whose Kimi K3 recently became the largest open-weight model release to date. As of publication, Ant Group had not published independent benchmark tables, license terms, or a model card.

The release adds to a wave of Chinese AI labs competing on efficiency rather than raw scale — a shift Nvidia’s Jensen Huang recently pressed dozens of companies to support publicly, arguing open models are essential to keeping pace in the US-China AI race.