📊 Benchmarks — post your Mac's local-LLM tok/s

[kennss]

Help build a crowd-sourced local-LLM speed table for Apple Silicon.

SiliconScope can measure any loaded model's decode speed and energy efficiency — the "Measure tok/s" button in the AI Runtime card (or sscope-cli --bench). Posting your numbers here lets others see what to expect from each chip before they buy.

Please reply with:

• Chip — e.g. M1 Max (10C / 32-GPU) · M4 Pro · M3 Ultra
• Unified memory — e.g. 64 GB
• Model + quant — e.g. gemma 26B (MoE) Q4_K_M
• Runtime — Ollama / LM Studio / llama.cpp
• tok/s — decode rate
• tok/Wh — efficiency (optional)

Example:

M1 Max · 64 GB · gemma 26B (MoE) Q4_K_M · Ollama · 59 tok/s · 4608 tok/Wh

Every data point helps someone pick the right Mac. 🙏