llama.cpp makes CPU-only codegen viable

This is not a product announcement, but it is exactly the kind of field-tested local inference guidance AI developers actually use when deciding whether old hardware is worth salvaging.

• –The strongest takeaway is that RAM capacity and CPU throughput can matter more than weak consumer GPUs for batch-style local codegen
• –Qwen3.5-27B Q4 emerges as a realistic target size for a 32GB CPU box, which is useful guidance for anyone planning overnight or queued jobs
• –llama.cpp server mode is the practical enabler here because sequential request handling turns slow token generation into a workable automation pipeline
• –The thread also reinforces a common local-LLM lesson: 4GB VRAM is usually too constrained to be worth designing around for modern coding models