Ant Group has released Ling 3.0 Tiny, which scores 25 on the Artificial Analysis

Ling 3.0 Tiny is small in parameters but still writes a lot, by Artificial Analysis's measurements. The open-weights reasoning model, released by Ant Group through its Ant Ling account, is a mixture-of-experts model: of its 7.9 billion parameters, only 1.3 billion are active for each token. It scores 25 on the Artificial Analysis Intelligence Index, comparable to gpt-oss-120b (high, 24) with 15 times fewer total and 4 times fewer active parameters. That places it on the Pareto frontier for intelligence against active parameters, meaning no open-weights model with as few active parameters scores higher, and Artificial Analysis says it is small enough to run locally in many settings.

The small size does not cut its output. Ling 3.0 Tiny generated 213 million output tokens to run the Intelligence Index, nearly as many as the much larger Ling 3.0 Flash (240 million), which has 124 billion total and 5.1 billion active parameters and scores 38. Artificial Analysis presents that high token usage as the trade-off for its parameter efficiency.

Its knowledge gains come from declining questions rather than answering more of them. On AA-Omniscience it scores -19, 59 points better than Ling-mini-2.0, mainly because its hallucination rate fell from 96% to 30% while accuracy stayed at 9%; it attempted only 37% of the questions instead of guessing. Artificial Analysis calls that accuracy relatively low among similarly sized open models. In agent tests it scores 21% on τ³-Banking, ahead of Qwen3.6 27B (17%), but reaches only 772 Elo on GDPval-AA v2, behind NVIDIA's Nemotron 3.5 Lightning (824), which it slightly outscores on the overall index. The weights are on Hugging Face under the MIT license, and the model is served through InclusionAI's first-party API and Novita.