{
  "name": "llama.cpp",
  "icon": "https://bb10repo.whytofly.de/img/bb10-apkui.png",
  "category": "Utilities",
  "type": "cli",
  "homepage": "https://gitacle.strangled.net/whytofly/llama.cpp_bb10",
  "description": "llama.cpp cross-compiled for BB10 (QNX 8 / Cortex-A9 / NEON). Runs small quantized LLMs (TinyLlama-1.1B Q2_K, SmolLM2-135M) locally on the device — no network, no cloud. Statically linked so no runtime libraries to ship.\n\nIncludes:\n- llama-completion : chat / text generation (formerly llama-cli)\n- llama-bench     : token-throughput benchmark\n- llama-tokenize  : tokenize text against a model\n- llama-quantize  : convert F16 / Q8 models down to Q2_K / Q4_K\n\nRun from Term49: `llama-completion -m /accounts/1000/shared/documents/your-model.gguf -p \"Hello\" -n 64 -t 2`. Expect tens of seconds per token on Cortex-A9 — fine for novelty / offline use, not chat. Source + build notes at https://gitacle.strangled.net/whytofly/llama.cpp_bb10.",
  "history": [
    {
      "version": "0.1",
      "released": "2026-06-15",
      "changes": "- Initial BB10/QNX cross-compile based on upstream llama.cpp 0.15.1 (commit 6e9007a).\n- Built against the bb10-gcc9 QNX 8 toolchain (Cortex-A9, NEON, VFPv3, soft-float ABI).\n- Static libgcc + libstdc++ so the .apk has zero runtime dependencies beyond stock BB10 libc/libsocket/libm.\n- GGML_LLAMAFILE=OFF because its SGEMM kernel needs hardware FP16 (Cortex-A15+); a hand-written NEON software fallback would recover some perf if needed."
    }
  ]
}