Just now, Ant stood on Kimi’s shoulders and open sourced a new model

Zhidongxi Author | Li Shuiqing Editor | Xinyuan Zhidongxi reported on August 11 that today, Ant Bailing Big Model open sourced a lightweight hybrid inference MoE model-Ling-3.0-tiny. The total parameter amount of this model is 7.9B, and the amount of activated parameters during inference is 1.3B. It is mainly designed for efficient local deployment. It provides three versions of BF16, FP8 and INT4 simultaneously, and has been verified on DGX Spark, MacBook, Mac mini and other devices, taking into account performance, efficiency and deployability. In the Artificial Analysis Intelligence Index evaluation system, Ling-3.0-tiny scored 25 points, only 1 point lower than Gemma-4-26B-A4B, and higher than gpt-oss-120B (high), Qwen3.5-9B, Gemma-4-12B and Gemma-4-E4B. The key points of the model are summarized as follows: 1. Efficient hybrid linear architecture: Ling-3.0-tiny uses a 3:1 alternating stacking structure of KDA (Kimi Incremental Attention) self-developed by Dark Side of the Moon and MLA (Multiple Latent Attention) self-developed by DeepSeek, and is equipped with a sparse MoE feedforward network composed of 128 routing experts to build an efficient hybrid linear architecture. Each token only activates 8 routing experts and 1 shared expert, allowing the model to achieve a balance between long context modeling capabilities, parameter efficiency and computational cost. 2. Native hybrid reasoning and agent capabilities: Ling-3.0-tiny takes into account both fast response and multi-step reasoning capabilities. Thinking mode can be flexibly turned on or off in a single request through the enable_thinking parameter. The model has balanced performance in general agent tasks, code generation, mathematical science reasoning, and instruction following scenarios. 3. Local and edge deployment: Under FP8 accuracy, Ling-