Skip to main content

AntBing's Ling-3.0-Flash: Small Model, Big Brain

AntBing has officially launched Ling-3.0-Flash, a next-generation native hybrid inference model that punches well above its weight. Released on July 24, this model packs 124 billion total parameters but only activates 5.1 billion per computation—a design choice that slashes computing costs without sacrificing performance.

In benchmarks covering basic reasoning, instruction following, and long text processing, Ling-3.0-Flash holds its own against—and sometimes beats—industry leaders with two to three times its parameter count. That's what the team calls a higher "intelligence efficiency ratio." For businesses, it means more bang for the buck.

Image

But here's where it gets interesting: Ling-3.0-Flash was built with real-world agents in mind. The team trained it on over 10,000 real interaction scenarios, fine-tuning its ability to self-correct and plan long-term tasks. Whether it's writing code, breaking down complex projects, or digging through multiple sources for research, this model is designed to deliver end-to-end without going off the rails—a common headache with traditional models.

So how did AntBing pull off "small size, high intelligence"? They ditched the brute-force approach of just adding more parameters. Instead, they rethought the underlying computing architecture from the ground up. Ling-3.0-Flash uses a hybrid attention design that alternates KDA linear attention layers and MLA layers in a 5:1 ratio. This mix balances long-context efficiency with raw capability.

Image

Another upgrade: the previous Lightning Attention has been replaced with KDA (Kimi Delta Attention). This new mechanism introduces fine-grained diagonal gating in the Delta Rule state update, helping the model remember key details when processing long documents or code repositories. Plus, they compressed the expert activation ratio per token from 1/32 to 1/64, squeezing out even more efficiency.

To make AI agents run faster and more reliably, AntBing also built a supporting engineering architecture. A hierarchical caching system at the cluster level cuts repeated calculations in long conversations, reducing first-character response latency by 60% to 80% for long inputs. And an upgraded multi-agent collaboration setup lets different agents work together and double-check each other, slashing the risk of errors in high-frequency online services.

Key Points:

  • Ling-3.0-Flash: 124B total params, 5.1B activated per task.
  • Matches or beats larger models in reasoning, instruction following, and long text.
  • Trained on 10,000+ real agent scenarios for better task completion.
  • Hybrid attention architecture (KDA + MLA) boosts efficiency.
  • Free on OpenRouter for one week, then open-sourced.