開放模型與推論系統
Ling-3.0-flash Uses 5:1 Hybrid Attention and 1/64-Sparse MoE to Reduce Activated Parameters per Token to 5.1B
InclusionAI has released Ling-3.0-flash, a 124B-parameter model that uses KDA, MLA, and a highly sparse MoE architecture for long-horizon agentic reasoning. The API is now available, but the weights are not scheduled for release until August, so its performance claims cannot yet be fully reproduced.
InclusionAI, an Ant Group subsidiary, has released Ling-3.0-flash. The model has 124B total parameters but activates only about 5.1B for each token. Rather than simply scaling down an existing Transformer, its architecture alternates KDA linear attention and MLA layers at a 5:1 ratio, combined with fine-grained diagonal gating and a 1/64-sparse MoE. The design aims to balance state memory, full-attention quality, and inference cost while offering both thinking and non-thinking modes.
According to the company, Ling-3.0-flash matches or outperforms the previous-generation Ring-2.6-1T on most internal benchmarks, despite having roughly one-eighth as many total parameters and one-twelfth as many activated parameters. The model targets coding, general computer-use tasks, and deep-research agents, and was trained using more than 10,000 interactive environments. Parameter efficiency, however, does not necessarily translate into real-world serving speed. MoE routing, inter-GPU communication, KV-state management, and batch size can all prevent theoretical FLOPs savings from scaling linearly into higher throughput.
The inference stack also integrates SGLang HiCache and Mooncake hierarchical caching, using separate prefill and decode pools together with cluster-shared L3 cache to avoid repeatedly computing shared prefixes during long-running interactions. The team says this can reduce time to first token for long inputs by 60% to more than 80%, but it has not yet disclosed complete details about the hardware, workload shapes, cache hit rates, or comparison settings. The model is currently available for testing through APIs such as OpenRouter, while its weights are scheduled to be released after the free API promotion ends. Engineering teams should next watch for the final license, native vLLM and SGLang support, single-node VRAM requirements, and whether the released weights can reproduce the reported long-context and agent benchmarks.