Skip to main content

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

AntBing has officially launched Ling-3.0-Flash, a next-generation native hybrid inference model that turns the 'bigger is better' assumption on its head. Released on July 24, this model boasts 124 billion parameters total, but here's the kicker: it only activates 5.1 billion per computation. That's like having a massive library but only pulling out the exact book you need—saving energy and time.

Despite its lean design, Ling-3.0-Flash holds its own against industry giants with two to three times its parameter count. In benchmarks for reasoning, instruction following, and long-text processing, it doesn't just keep up—it often surpasses them. The result? A higher 'intelligence efficiency ratio' that makes it both smarter and cheaper to run.

Image

Built for Real-World Agents

What sets Ling-3.0-Flash apart is its focus on practical agent applications. 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 researching across multiple sources, the model delivers more reliably than traditional models that tend to go off the rails on big jobs.

The Secret Sauce: Hybrid Attention

So how did they pack so much intelligence into a small package? They ditched the 'just add more parameters' approach. Instead, Ling-3.0-Flash uses a hybrid attention architecture from the get-go. It alternates KDA linear attention layers with MLA layers in a 5:1 ratio, striking a sweet spot between long-context efficiency and raw capability.

They also upgraded the previous Lightning Attention to something called 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 wading through long documents or code repositories. Plus, they squeezed the expert activation ratio per token from 1/32 down to 1/64, getting more bang for each computational buck.

Image

Speed and Stability for Real Apps

To make AI agents run faster and more reliably, AntBing built a supporting engineering architecture. A hierarchical caching system at the cluster level cuts repeated calculations in long conversations, slashing 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, reducing the risk of errors in high-frequency online services.

Ling-3.0-Flash is now available on OpenRouter for free for one week. After that, it will be fully open-sourced—a move that could shake up the AI landscape by proving that sometimes, less really is more.

Key Points

  • Model: Ling-3.0-Flash, 124B total parameters, 5.1B activated per computation.
  • Performance: Matches or exceeds models 2-3x its size in reasoning, instruction following, and long-text tasks.
  • Architecture: Hybrid attention with KDA and MLA layers (5:1 ratio), plus KDA attention with diagonal gating.
  • Efficiency: Expert activation ratio reduced to 1/64 per token.
  • Availability: Free on OpenRouter for one week, then open-sourced.
  • Real-world focus: Trained on 10,000+ agent scenarios, with caching and multi-agent collaboration for speed and stability.