Ant Ling releases Ling-3.0-flash: 124B total, 5.1B active params, matching 1T-class flagship performance

Ant Ling's Ling-3.0-flash uses 124B total and 5.1B active parameters to match or beat its previous 1T-class Ring-2.6-1T on reasoning, instruction following, and agent tasks. The key change is a native hybrid linear attention architecture stacking KDA and MLA layers at a 5:1 ratio, with expert activation dropped to 1/64, trading parameter scale for higher intelligence density. On the agent side, over 10,000 interactive training environments were added, achieving closed-loop execution for Coding, General, and Deep Research agents; it hits 25.3% on MiniAppBench, on par with open-source SOTA models twice its size. For inference, SGLang HiCache + Mooncake hierarchical caching cuts TTFT by 60–80%+ on long-context inputs. The post does not disclose API pricing or availability date.

Why it matters: Ant Ling drops Ling-3.0-flash: 124B total, 5.1B active params, matching their previous 1T flagship on reasoning and agent tasks. Three concrete architecture improvements, not fluff. Held back from 85+ because it's a single-source release with no third-party benchmarks yet, and...