Ant Group's Bailing Open-Sources Lightweight Hybrid Model Ling-3.0-tiny

Ant Group's Bailing team open-sourced the lightweight hybrid model Ling-3.0-tiny on Hugging Face on August 11. The model uses a MoE architecture with 7.9B total parameters and activates 1.3B per token. It comes in BF16, FP8, and INT4 versions to fit different deployment needs.

Ling-3.0-tiny focuses on low-cost inference. It uses an efficient hybrid linear design, stacking KDA and MLA in a 3:1 ratio, plus a sparse MoE feed-forward network with 128 routed experts. This balances context handling and computing cost.

The model supports fast response and multi-step reasoning, designed for local use. It has been tested on devices like NVIDIA DGX Spark and Apple Silicon MacBooks. On a MacBook with M4 Pro, it reaches about 86-90 tokens per second, using around 8.34 GiB peak memory at 8K context.