Oracle69digitalmarketing/adtc-2026-laptop-llm

Africa Deep Tech Challenge 2026 — offline on-device AI agent optimized for commodity African laptops. Offline Coding Assistant for African Laptops «ADTC 2026 · Coding Assistants · Offline / Edge AI» A reproducible local AI coding-assistant configuration designed for constrained hardware: approximately 8 GB laptops, CPU-only inference, and environments where reliable cloud connectivity cannot be assumed. The project combines a compact instruction-tuned GGUF model with llama.cpp to provide local coding assistance without sending prompts or source code to a cloud inference API. One compact model. Multiple device classes. Local inference. --- Why This Project AI coding assistants are increasingly useful, but many students and developers work with: - Modest laptops with limited RAM - CPU-only hardware and no discrete GPU - Unreliable or expensive internet connectivity - Privacy requirements that make cloud inference undesirable - Limited access to high-performance computing hardware This project explores a practical alternative: «Put the model on the device and run the assistant locally.» The primary competition target is an approximately 8 GB CPU-only laptop. We also validated the same compact model on an ARM64 Android phone through Termux. That mobile validation is supplementary to the ADTC submission, but demonstrates that the deployment can extend beyond the laptop. --- What We Built We built a compact offline coding-assistant configuration around: - SmolLM2-135M-Instruct - GGUF Q4_K_M quantization - llama.cpp - CPU-only inference - The official ADTC participant profiler - Reproducible model download and benchmark tooling The core workflow does not require a cloud inference API. Once the model and runtime are available locally, prompts and source code can remain on the device. --- Final Model Component| Configuration Model| SmolLM2-135M-Instruct Parameters| ~135M Quantization| GGUF Q4_K_M Model size| ~100 MiB Runtime| llama.cpp Primary target| ~8 GB CPU-only laptops Additional validation| ARM64 Android + Termux Inference| CPU-only The model is …