b10506

ggml-org/llama.cppb10506Aug 20, 2026by github-actions[bot]same day as b10505

What changed since b10505

low risk

b10506 adds a Metal-only optimization that dequantizes q8_0 weights using packed types, with no other changes noted.

Improvements

  • Metal backend now dequantizes q8_0 tensors using packed types, improving Apple Silicon throughput.
  • The change is scoped to the Metal backend, so CPU, CUDA, Vulkan, and other backends are unaffected.

Upgrade notes

This is a single-commit increment over b10505, so no config or migration work is needed. macOS/iOS users on Apple Silicon should see the benefit automatically; other platforms can upgrade with no expected behavior change. If you pin builds, note that the previous release's dedup-cache-models server preset remains available.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Optimizes the macOS Metal backend by implementing q8_0 dequantization using packed types, which should improve performance on Apple Silicon.

Key Highlights

  • Dequantize q8_0 using packed types on Metal
  • Optimize Metal backend performance

New Features

  • Metal backend optimization for q8_0 dequantization

Release Notes (excerpt)

<details open> metal : dequantize q8_0 using packed types (#27370) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.14)DISABLED - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories