nile-agi/LMc

Edge-first AI inference for everyone. Pure C99 · No dependencies · Runs on any CPU · Built for Africa and the Global South. Local Model computing ← Local Machine learning Models computation on your low-end edge device ➝ by custom research --- --- ᯓ➤ 📨 Email Us ᯓ➤ 🌐. Visit ᯓ➤ [[in] lınkedln](linkedin.com) --- LMc is an AI inference engine written in pure C99. It runs machine learning models locally on **any device** — no cloud, no GPU required. GGUF-native. Zero external dependencies. Runs on x86, ARM, Android, Windows. LMc is designed for the reality of computing in Africa and the Global South: low-spec phones, aging laptops, shared computers, and limited bandwidth. Where llama.cpp is a toolkit, LMc is a standard — the **FFmpeg of AI inference**. **Current status:** Proof of concept — GPT-2 124M working end-to-end. Architecture and extension points are production-ready. New models and hardware backends plug in cleanly. --- ``` ./lmc --model models/gpt2-xl.gguf \ --prompt "The meaning of life is" \ --n-predict 128 --temp 0.7 --threads 4 ``` --- ## Features - **GGUF-native** — reads models directly from the llama.cpp / HuggingFace GGUF format - **All GPT-2 variants** — Small (124M), Medium (345M), Large (774M), XL (1.5B) - **Rich quantisation support** — F32, F16, Q2_K, Q3_K, Q4_0, Q4_1, Q4_K, Q5_0, Q5_1, Q5_K, Q6_K, Q8_0, IQ3_XXS, IQ3_S, IQ4_XS - **Optimised kernels** — 16-wide matmul unroll, head-major KV cache, fast GELU, OpenMP parallelism - **Edge-friendly** — Raspberry Pi 4 at 8 t/s (GPT-2 Small Q4_K_M, 4 threads) - **Portable** — Pure C99, no BLAS, no external libs beyond `-lm` --- ## Quick Start ### 1. Build ```bash make # single-threaded make omp # OpenMP multi-threaded (recommended) make help # all targets ``` macOS requires `brew install libomp` for OpenMP. Windows: use MSYS2/MinGW64 terminal. ### 2. Download a model (GGUF) There are a lot of available source you can use to get the weight from, I find the following to be simple sources. You will probably find a lot of models version (I me …