Ant Group Releases Compact Ling 3.0 Tiny Model
The 7.9B-parameter mixture-of-experts model targets economical local and agent deployment, activating only 1.3B parameters per token.
A small model built around sparse activation
Ant Group’s InclusionAI team has released Ling 3.0 Tiny, a compact open-weights language model with 7.9 billion total parameters and 1.3 billion activated for each token. The sparse mixture-of-experts design is intended to provide more capability than the active parameter count would normally suggest, while keeping inference costs and memory traffic suitable for constrained deployments.
Artificial Analysis gave the model a score of 25 on version 4.1.1 of its Intelligence Index. That places Ling 3.0 Tiny among the stronger small open-weight systems tracked by the evaluator, although it remains well below frontier-scale models in absolute capability. Its score of minus 19 on the evaluator’s Omniscience Index—a measure designed to penalize confident fabrication—was described as relatively competitive for its size.
The efficiency picture is less straightforward. Artificial Analysis found that Ling 3.0 Tiny generated about 51,000 output tokens per Intelligence Index task. That exceeded the roughly 36,000 used by the larger Ling 3.0 Flash and was approximately 65% more than efficient competitors including Nemotron 3.5 Lightning and comparable Qwen models. A low active-parameter count can reduce the cost of each generated token, but excessive reasoning length may return some of those savings through longer runtimes and higher aggregate compute consumption.
Why it matters
Ling 3.0 Tiny adds another credible Chinese model to the increasingly competitive sub-10-billion-parameter segment. Such models matter less as chatbot flagships than as components for local assistants, high-volume agents and specialized systems where latency, privacy and operating cost outweigh maximum benchmark performance. Its unusually long reasoning traces also underline why buyers should evaluate complete task economics—not parameter counts or per-token prices alone. Independent tests of latency, memory use and task-level cost across deployment hardware will determine whether its sparse architecture translates into a practical advantage.