b10414

ggml-org/llama.cppb10414Aug 13, 2026by github-actions[bot]same day as b10413

What changed since b10413

low risk

b10414 adds TQ2_0 ternary quantization support to the Metal backend and optimizes the mul_mv kernel.

New features

  • Metal backend now supports the GGML_TYPE_TQ2_0 ternary 2-bit quantization type, enabling ternary-quantized models to run on Apple GPUs.

Improvements

  • mul_mv kernel on Metal reworked to use float ops over integer ops and precomputed sums.
  • Loop coefficient hoisting and contiguous y loads reduce overhead in the Metal mul_mv path.

Upgrade notes

This is a routine incremental build with no config or API changes; existing flags and model files continue to work. Apple Silicon users working with TQ2_0 ternary quants should see these models become usable and faster, while other backends are unaffected.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

Adds support for the TQ2_0 (ternary 2-bit) quantization type to the Metal backend and optimizes the mul_mv kernel.

Key Highlights

  • Add TQ2_0 support to Metal backend (#26980)
  • Optimize mul_mv kernel for Metal

Release Notes (excerpt)

<details open> metal : add TQ2_0 support (#26980) * metal: add TQ2_0 support Add support for the GGML_TYPE_TQ2_0 (ternary, 2 bits per element) type in the Metal backend. Assisted-by: llama.cpp:DeepSeek-v4-Flash-0731 * cont : optimize mul_mv kernel - float ops over integer ops - precalculate sums - hoist coef out of the inner loop - contiguous y loads llama.cpp:DeepSeek-v4-Flash-0731 </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories