Ant's Tiny New MoE Beats Qwen3.5-9B on Agent Benchmarks

Ant Bailing (Inclusion AI), Ant Group's AI lab, has released Ling-3.0-tiny — a small mixture-of-experts model built for local deployment and agent use cases. It has 7.9 billion total parameters, but only activates 1.3 billion per token. It also supports two modes, a reasoning mode and a quick-answer mode, and can call tools to complete tasks.

Even with just a slice of its weights active at a time, it's already beating bigger models on some agent benchmarks. In Ant's own tests, Ling-3.0-tiny outscored Qwen3.5-9B on both the GDPval v2-AA and TAU3-Banking-AA general-purpose agent benchmarks.

Ling-3.0-tiny is live on OpenRouter and Vercel AI Gateway, free to try, and the model weights will be open-sourced later.