Running a 35B Model on a Low-Spec CPU Laptop to Generate Games
A user demonstrated running a 35-billion parameter model (Qwen3.6 35B quantized to IQ2_XXS 2-bit format) on a budget Intel i3 laptop with 8GB RAM and zero dedicated VRAM. The setup successfully generated a functional single-file HTML/JS 2D RPG in 24 minutes at a speed of 2 tokens per second using llama.cpp. This demonstration highlights how extreme model quantization enables complex AI tasks on standard consumer hardware without requiring expensive GPUs. It makes running large language models locally far more accessible for developers and hobbyists operating on low budgets. The model was loaded using llama-server with CPU-only processing (`-ngl 0`), an 8-bit KV cache, and an 8,192-token context length. Disabling the reasoning mode prevented the model from wasting computation time on excess thinking tokens, ensuring steady output on CPU.
## BACKGROUND
IQ2_XXS is an ultra-low 2-bit quantization technique designed to compress large models so they fit in limited memory while preserving functional intelligence. GGUF is a single-file binary format designed for efficient memory loading and CPU inference within the llama.cpp ecosystem.