Ant Group Releases 7.9B Ling-3.0-tiny with 90 Token/s Local Speed

Ant Group opens up Ling-3.0-tiny weights under MIT license. The 7.9B MoE model activates only 1.3B parameters per token, enabling high-speed local inference on consumer hardware.

Woofun AI reports that Ant Group has released the open-source weights for Ling-3.0-tiny, providing BF16, FP8, and INT4 versions under the MIT license. The 7.9 billion parameter model utilizes a mixture-of-experts architecture with 128 routing experts, activating only 1.3 billion parameters per token alongside one shared expert. It employs a hybrid attention mechanism combining Kimi Delta Attention and MLA in a 3:1 layer ratio. Performance benchmarks indicate the FP8 variant delivers 86–90 tokens per second on an M4 Pro MacBook and 100–105 tokens per second on a DGX Spark.