FULL STORY

AntLingAGI Ling-3.0-tiny: From Launch to Open Source

AntLingAGI released and open-sourced Ling-3.0-tiny, a lightweight hybrid reasoning model. It quickly topped the Hugging Face trending list due to its efficient sparse activation.

2026-08-07 ~ 2026-08-12 · 2 episodes · 9 posts

Episode 1 · AntLingAGI Releases Ling-3.0-tiny Native Hybrid Reasoning Model (2026-08-07, 2 posts)

AntLingAGI has released Ling-3.0-tiny, a native hybrid reasoning model with 7.9B total parameters that activates only 1.3B per token. It is designed for real-world tasks, math, instruction following, and resource-sensitive deployments.

Episode 2 · Ant Group Open-Sources Ling-3.0-tiny: 1.3B Active Parameters, Top HF Trending (2026-08-11, 7 posts)

Following Ling-3.0-flash, Ant Group's Bailing team open-sourced Ling-3.0-tiny, a lightweight hybrid-inference MoE model with 7.9B total parameters and only 1.3B active parameters per token, designed for extreme inference efficiency and real-world task execution. It has already reached the Hugging Face trending list, offering a low-compute option for local agent deployment.

Confirmed

  • Architecture & Parameters: MoE with 128 experts, 7.9B total, 1.3B active. Uses a 3:1 KDA (Kimi Incremental Attention) and MLA architecture, supporting 262K context.
  • Open-source specs: Available in BF16, FP8, and INT4 precisions, includes custom code, MIT license.
  • Performance: Scored 16.7 on Artificial Analysis intelligence index. Tests show overall performance between 4B and 12B Qwen models, and with 1.3B active parameters it outperforms some larger models (e.g., 31B or GPT-OSS 120B on some metrics).
  • Use cases: Designed for reasoning, tool calling, and agent tasks, suitable for efficient local deployment.

Why it matters

Ling-3.0-tiny achieves above-class performance with extremely low active parameters, significantly lowering the compute barrier for edge or local agent deployment, providing a competitive open-source option for building efficient real-world task execution systems.