Old i7 setup exposes CPU LLM speed limits

The practical takeaway is that old CPU rigs can still be useful for local inference, but memory bandwidth and model choice dominate throughput more than raw CPU age.

• –Multiple commenters reported expectations around roughly 2-8 tok/s for 12B-14B-class models on CPU-only setups, making 12 tok/s optimistic for this hardware tier.
• –The thread repeatedly points to RAM bandwidth and dual-channel configuration as key constraints, which matches broader llama.cpp discussions about token generation being memory-bound.
• –MoE models (for example, Qwen 3.5 35B with low active parameters) were recommended as a way to improve perceived speed on limited hardware.
• –llama.cpp tuning (threads, batch/context settings, quantization level) can materially change results, so first-run defaults should be treated as a baseline, not a final verdict.
• –For hobby or background/agentic workloads, commenters framed this class of machine as “good enough” if expectations are set around latency.