llama.cpp Release b11279 Adds Support for GLM-5.3-Flash Architecture
Open-source LLM inference engine llama.cpp has released build b11279, adding native architectural support for Z.ai's GLM-5.3-Flash (GLM5-Next) model. The update introduces memory hybrid index integration, multi-stream handling, persistent k-pool layouts across micro-batches, and optimizations for long-context decoding. Adding support for GLM-5.3-Flash enables developers and researchers to execute Z.ai's latest highly efficient hybrid model locally on consumer devices. It expands the range of state-of-the-art open architectures accessible via lightweight CPU and GPU acceleration frameworks. The pull request (#27773) integrates recurrent state rollback checkpoints, refactors memory graph base handling, and improves compute buffer allocations to speed up prefill operations. Additionally, it addresses graph allocation issues across CUDA, ROCm, Vulkan, Metal, and WebGPU backends.
## BACKGROUND
llama.cpp is an open-source C/C++ library designed to run large language models locally with high efficiency across various hardware backends. GLM-5.3-Flash is a native multimodal LLM developed by Z.ai that features a redesigned architecture tailored for high capability and low inference cost.