Ant Group Releases Ling-3.0 Small Model, Outperforms Qwen3.5-9B in Two Agent Evaluations

According to monitoring by Beating, Ant Group's Inclusion AI has launched Ling-3.0-tiny. This is a small MoE model designed for local deployment and Agent scenarios, featuring 7.9 billion parameters, with only 1.3 billion parameters activated per token. It supports both reasoning and quick response modes and can also utilize tools to execute tasks. Despite activating only 1.3 billion parameters at a time, it has surpassed the larger Qwen3.5-9B in certain Agent tasks. In official tests, Ling-3.0-tiny achieved higher scores in two general Agent evaluations: GDPval v2-AA and TAU3-Banking-AA. Ling-3.0-tiny is now available on OpenRouter and Vercel AI Gateway, and it is currently free to try. The model weights will also be open-sourced in the future. [Original Link]