Chinese and English notes from building a local inference runtime.

I write about measurements, false starts, and the hypotheses the numbers killed. A result without its script, its error bars, and its failed alternatives is a claim, not a result.

• Flint Inference on Apple Silicon: rec Megakernel and Tempo9 Compared
  2026-10-27 · English · Comparing rec and Tempo9 on one M5 Pro: single-stream decoding, concurrency, prefill, cross-session prefix caching, CPU overhead, and megakernel tradeoffs.
• Run Gemma 4 on Your Mac with Tempo9
  2026-10-27 · English · Install Tempo9, download Gemma 4 E2B or E4B GGUF, run a local OpenAI-compatible API, and see a short M5 Pro benchmark against llama.cpp.