My local AI 70% faster with llama.cpp (RTX 5060 8 GB, new MoE cache)

Описание: Same graphics card, same model, and yet 70% faster. I test llama.cpp's new option (November 7, 2026): the MoE expert cache with CPU share.

🖥️ Setup: RTX 5060 8 GB (PCIe 3), Xeon E-2146G, 32 GB RAM, Qwen3.6-35B-A3B Q4_K_M, 32k context

📊 Results (tokens/s):
• LM Studio: 33.1
• llama.cpp without cache: 30.2
• Small cache without CPU share: 24.8 (the trap)
• 3.8 GB cache only: 39.1
• 3.8 GB cache + CPU share 0.9: 51.3

⚙️ Flags: --cpu-moe --moe-cache-mib 3840 --moe-cache-cpu-share 0.9 (values found by moe-cache-tune.py)
⚠️ On Windows, the tuning script crashes (SIGINT): one-line fix explained in the video.

🎙️ Voice and 🎵 music generated locally (VoiceStudio + ACE-Step).

How much do you gain on your machine? Tell me in the comments 👇