llama.cpp makes 6GB GPUs viable

This is the classic “good enough to be useful, not good enough to be luxurious” local AI laptop. The winning move is a lightweight `llama.cpp` stack, or a wrapper like LM Studio or Ollama, paired with realistic expectations about model size, context length, and speed.

• –Community guidance in the thread clusters around quantized 4B-9B models for daily use, with larger models only making sense if you are willing to spill heavily into system RAM and tolerate slow responses.
• –`llama.cpp` is the key enabler here because it can auto-detect hardware, choose quantization kernels, and offload only part of the model to GPU, which is exactly what a 6GB VRAM machine needs.
• –LM Studio’s Windows docs recommend as little as 4GB of dedicated VRAM and include GPU-offload and memory-estimate controls, making it a friendlier way to test what fits before dropping into CLI workflows.
• –For code-heavy work, small code-tuned options such as Qwen2.5-Coder 7B are still a solid baseline because they are built for code generation, repair, and reasoning while offering quantized GGUF variants that suit constrained hardware.
• –The Intel iGPU memory shown in Task Manager is mostly shared system RAM, not a seamless extra VRAM pool; Intel-specific backends like BigDL-LLM can target iGPU acceleration, but that is a separate and more finicky path, not a free boost in mainstream Windows local runners.