llama.cpp b10269 fixes dflash wo_a reshape loading bug
New llama.cpp release patches crash on model load for DeepSeek-style architectures
ggml-org's llama.cpp, the star-123k open-source library that powers countless local AI applications, has shipped a new patch release tagged b10269. The only functional change noted is a fix for "dflash wo_a reshape on load" (#26577). This addresses a bug where the weight matrix wo_a in DeepSeek-style flash-attention models was being reshaped incorrectly when loading a GGUF model, potentially causing crashes or corrupted inference. For developers who run Mixture-of-Experts models locally, this is a welcome stabilization fix.
The release includes a comprehensive set of prebuilt binaries across every major backend: Apple Silicon (including KleidiAI-accelerated builds), Intel x64, Vulkan, ROCm 7.2, OpenVINO, SYCL (FP16/FP32), CUDA 12.4/13.3, OpenCL for Adreno, and HIP for Windows, plus multiple openEuler variants. There are also dedicated iOS and Android packages. This release underscores llama.cpp's role as the universal runtime for on-prem and edge LLM inference. Users should upgrade to b10269 to avoid model-loading errors with affected architectures, especially when using deepseek-based GGUF files.
- Fixes dflash wo_a reshape crash on model load (#26577)
- Prebuilt binaries for CUDA 12/13, ROCm 7.2, Vulkan, OpenVINO, SYCL, and more
- Released Aug 4, 2025; supports Apple Silicon, Windows, Linux, Android, iOS
Why It Matters
Stability fix ensures smoother local AI inference for developers using DeepSeek-flash architectures on llama.cpp.