# llama.cpp adds Vulkan sparse flash attention for quantized keys and values

> The open‑source LLM runtime now supports a Vulkan‑based sparse flash‑attention mode for quantized K/V data.

Oossa · 2026-10-05 · https://oossa.com/en/llama-cpp-adds-vulkan-sparse-flash-attention-for-quantized-keys-and-values

The ggml‑org team released llama.cpp version b11413 on Oct 5 2026. The update adds a Vulkan implementation of sparse flash attention that works with quantized key/value tensors. The change is described as "sparse flash attention for quantized K/V" and is aimed at GPU‑accelerated inference. The release also lists many platform builds, from macOS Apple Silicon to Windows CUDA and Vulkan.

## The facts

- Release version b11413 published on Mon Oct 05 2026
- New feature: Vulkan sparse flash attention for quantized K/V

## Why it matters

Developers can now run quantized models faster on GPUs that support Vulkan, lowering compute cost for edge devices.

## Sources & references

1. [ggml-org/llama.cpp b11413](https://github.com/ggml-org/llama.cpp/releases/tag/b11413) – llama.cpp, 2026-10-05

Last updated: 2026-10-05