Ant Group Launches Ling-3.0-flash Hybrid Reasoning Model with 124B Parameters
Ant Group has launched Ling-3.0-flash, a native hybrid reasoning model with 124B total parameters and only 5.1B active parameters. Despite its smaller size, it matches or exceeds the performance of the previous 1T-parameter flagship model, Ring-2.6-1T. This release showcases a significant leap in LLM efficiency, proving that sparse MoE and hybrid attention architectures can deliver flagship-level performance at a fraction of the computational cost. It lowers the barrier for deploying fast, cost-effective, and complex AI Agent workflows in real-world production environments. The model utilizes a hybrid linear attention architecture stacking Kimi Delta Attention (KDA) and Multi-head Latent Attention (MLA), alongside a 1/64 sparse Mixture of Experts (MoE). Additionally, it integrates SGLang HiCache and Mooncake hierarchical caching, reducing Time-to-First-Token (TTFT) by 60% to over 80% for long context interactions.
## BACKGROUND
Traditional Transformer models suffer from high computational complexity when handling long sequences due to the quadratic scaling of standard attention mechanisms. To address this, linear attention mechanisms like Kimi Delta Attention (KDA) are designed to improve hardware efficiency. Furthermore, distributed KV caching systems like Mooncake and SGLang HiCache optimize inference by storing and sharing previously computed context data, preventing redundant calculations and speeding up response times.