Ant Group Open Sources Ling-3.0-tiny: 7.9 Billion Parameters, Achieves 90 Tokens/s on M4Pro Locally

According to the Dynamic Beating Monitor, AntBird has officially released Ling-3.0-tiny, providing three versions with BF16, FP8, and INT4, all under the MIT license on the Hugging Face page. The model has a total of 7.9 billion parameters, with only 1.3 billion activated per token, mainly targeting on-premises deployment and agent scenarios.

Ling-3.0-tiny features 128 routing experts, with each token selecting only 8 of them, plus 1 shared expert used by all tokens. The attention module adopts a 3:1 KDA to MLA hybrid architecture, meaning that for every 3 layers of Kimi Delta Attention, there is 1 layer of MLA.

The official statement indicates that the FP8 version can achieve around 86–90 Tokens/s on an M4 Pro MacBook and approximately 100–105 Tokens/s on a DGX Spark.