Integration Status of Ling-3.0-Flash Weights Across Major Inference Engines
The upcoming Ling-3.0-flash model, featuring a hybrid KDA and MLA attention mechanism, is seeing varied integration preparation across major LLM inference engines. SGLang has committed to day-0 support, vLLM plans to release support once weights are open-sourced, and llama.cpp currently lacks a conversion path for its Bailing MoE architecture. This highlights the coordination challenges between model creators and open-source inference frameworks during "announce-first, open-weights-later" releases. It also shows how specialized architectures like Bailing MoE can face delays in local deployment tools like llama.cpp due to a lack of community-driven conversion paths. While llama.cpp already supports the KDA kernels used in Ling-3.0-flash, the main blocker for GGUF support is the Bailing MoE conversion path, which remains unimplemented after previous pull requests for version 2.6 were closed. The open-source release of the weights is expected to follow the end of the model's free API access window on August 3.
## BACKGROUND
SGLang and vLLM are high-performance serving frameworks designed for deploying LLMs efficiently. Ling-3.0-flash utilizes a hybrid attention mechanism combining Kimi Delta Attention (KDA), a hardware-efficient linear attention module, and Multi-Head Latent Attention (MLA), which compresses the Key-Value (KV) cache to reduce memory usage. Mixture of Experts (MoE) is an architecture that activates only a subset of the model's parameters per token, and Bailing MoE is the specific variant used by Ant Group's Ling models.