llama.cpp - Termux (Tensor / Pixel)

Upstream llama.cpp, built for Android Termux (arm64) with bundled .so, tuned for Google Pixel / Tensor devices.

Prebuilt llama.cpp binaries for Android Termux (arm64) with bundled shared libraries for maximum CPU performance on Pixel/Tensor.

What This Is

• Prebuilt llama.cpp binaries for Termux on Android (arm64).
• Optimized build profile intended for Google Pixel / Tensor devices.
• This package bundles libllama/libggml* shared libs under lib/ and runs via thin wrappers.

Exact upstream commit + build flags for this specific release are recorded in docs/build_meta.txt.

Commands (No Clobbering Termux PKG)

This package intentionally installs namespaced commands so you can keep the Termux package llama-cpp installed for benchmarks:

• llama-cli-tensor
• llama-bench-tensor
• llama-server-tensor

Installation

pkg update && pkg upgrade -y
pkg install -y nodejs-lts openssl
npm install -g @mmmbuto/llama-cpp-termux-tensor

Automatic Runtime Deps (postinstall)

On Termux, the npm postinstall step will attempt to install missing runtime packages automatically via pkg:

• openssl (for libssl.so.3 / libcrypto.so.3)
• libc++ (for libc++_shared.so)

To skip (CI/dry runs):

export LLAMA_CPP_TERMUX_SKIP_PKG_INSTALL=1

Verify

llama-cli-tensor --version
llama-bench-tensor -h

Benchmarks (Pixel 9 Pro)

Bench parameters: threads=6 batch=256 ubatch=256 mmap=1 prompt=512 gen=256 reps=3

| Model (Q4_K_M) | PKG pp512 | TENSOR pp512 | PKG tg256 | TENSOR tg256 | |---|---:|---:|---:|---:| | Llama 3.2 3B | 10.46 | 25.42 | 4.33 | 4.82 | | Gemma3n E2B-it | 12.47 | 30.18 | 5.49 | 6.10 | | Phi 3.5 mini | 6.19 | 17.27 | 4.31 | 4.83 | | SmolLM2 1.7B | 16.62 | 48.79 | 6.98 | 7.88 | | Qwen3 1.7B | 12.85 | 40.59 | 6.55 | 8.05 |

These numbers are provided as a real-world example; your results will vary with temperature/throttling, Android build, and Termux version.

Build (Repro)

See docs/build_meta.txt for the exact build metadata.

High-level build steps:

pkg install -y git cmake ninja clang make openssl

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
git checkout 338085c69e486b7155e5b03d7b5087e02c0e2528

cmake -S . -B build-android-tensor -DCMAKE_BUILD_TYPE=Release \
  -DGGML_OPENMP=ON -DLLAMA_OPENSSL=ON \
  -DLLAMA_BUILD_SERVER=ON -DLLAMA_BUILD_TOOLS=OFF -DLLAMA_BUILD_TESTS=OFF \
  -DGGML_BACKEND_DL=OFF \
  -DGGML_NATIVE=ON

cmake --build build-android-tensor -j"$(nproc)"

Packaging Details (.so)

This package bundles these libraries under lib/:

• libggml*.so*, libllama*.so*, libmtmd*.so*

The wrappers in bin/ set LD_LIBRARY_PATH so the binaries use the packaged .so.

License

Upstream llama.cpp is MIT licensed. This package redistributes compiled binaries from upstream and includes the upstream LICENSE.