Tools · · 2 min read

llama.cpp b11716 runs Q1_0 and Q2_0 on Snapdragon Hexagon: HMX/HVX matmul paths, tested on SM8850

llama.cpp b11716 (October 17, 2026) adds Q1_0 and Q2_0 weight support to the Hexagon backend via PR #30122. Native HVX MUL_MAT, Q1_0/Q2_0 on the 2D HMX path, and 128/128 op tests passed on SM8850/v81. Here is what Snapdragon builders need to pull and how to select HTP0.

llama.cpp b11716, published October 17, 2026, is a Hexagon-focused pre-release: it adds Q1_0 and Q2_0 weight support on Snapdragon’s NPU path. If you already run GGUFs on Hexagon with Q4_0-class packs and want the newer 1-bit / 2-bit layouts, this is the build that unlocks them.

Builder question this answers: can stock llama.cpp decode Q1_0 / Q2_0 GGUFs on Hexagon HTP, and what do I need to install?

What PR #30122 adds

The only change called out in the b11716 release notes is PR #30122 (“hexagon: add support for Q1_0 and Q2_0”), co-authored with Qualcomm’s Max Krasnyansky and merged October 17, 2026. From the PR:

  • Native Q1_0 and Q2_0 MUL_MAT through HVX
  • Direct {-1,+1} expansion against tiled Q8_0 activations
  • Workers for 32×1, 32×2, 2D, 4D, and matvec
  • Exact host repacking and arbitrary byte-range readback
  • Q1_0 and Q2_0 weights enabled on the 2D HMX MUL_MAT path
  • Tiling fixes so MUL_MAT_NX no longer assumes square tiles (Q1/Q2 use non-square tiles)

Verification claimed in the PR: v73, v75, v79, and v81 builds passed; tested on SM8850 / v81 Android; Q1_0 128/128 and Q2_0 128/128 op tests passed. GET_ROWS support is explicitly deferred to a follow-on.

How to get a Hexagon-capable binary

b11716 ships prebuilt Snapdragon archives on the release page:

  • Linux arm64 Snapdragon: llama-b11716-bin-linux-arm64-snapdragon.tar.gz
  • Android arm64 Snapdragon: llama-b11716-bin-android-arm64-snapdragon.tar.gz

Official setup is in the Snapdragon backend README: cross-compile with the Qualcomm Snapdragon Toolchain images (ghcr.io/snapdragon-toolchain/arm64-android:v0.7 or arm64-linux:v0.7), or use ./scripts/snapdragon/build.py --target adb|linux …. CMake presets turn on GGML_HEXAGON=ON and GGML_OPENCL=ON.

On device, Hexagon shows up as an HTP “GPU-like” device. The docs’ default pattern is --device HTP0 (or HTP0:0) with -ngl 99. Example from the README:

./scripts/snapdragon/run.py --target adb --devices HTP0 -- \
  llama-cli -m Llama-3.2-1B-Instruct-Q4_0.gguf \
  -p "what is the most popular cookie in the world?"

Swap in a Q1_0 or Q2_0 GGUF once you have b11716+ Snapdragon binaries. For larger models, the same README documents layer-split across virtual sessions (HTP0:0,HTP0:1,…) and tensor-split across physical NPUs.

What this is not

  • It is not Prism ML’s ternary PQ2_0 / PTQ1_0 types — those remain outside upstream Hexagon until separately implemented. Upstream Q2_0 is the format this PR targets.
  • Hexagon support is still labeled experimental in the Snapdragon docs.
  • Pre-b11716 Snapdragon builds that already ran Q4_0 do not suddenly gain Q1_0/Q2_0 kernels; pull b11716 or newer.

If you prefer Qualcomm’s higher-level OpenAI-compatible wrapper instead of raw llama.cpp, see our GenieX on Snapdragon note. For the previous day’s CUDA Q2_K/Q3_K decode work, see llama.cpp b11713, and for the Discover Models UI in b11714.

Sources

Source: llama.cpp release b11716