Writing
Benchmarks, local models, and agents doing real work.
2026
6 postsWhich qwen Actually Fits a 16GB GPU
No qwen3.8 tag fits a 16GB card. Two Modelfile lines make it 79% faster than the default, and qwen3.5 9B with MTP runs Q8 weights at Q4 speed.
The Best gemma4 12B Quant for a 16GB Blackwell GPU
Benchmarking gemma4 12B quants on the cheapest 16GB Blackwell card. QAT beats Q4_K_M outright, and its FP4, FP8, and MTP speedups are all macOS-only on Ollama.
I Let My AI Agent Plan My Tokyo Marathon Trip
I gave an OpenClaw agent its own Gmail, iCloud, and browser, then let it plan eleven days in Tokyo and Hakuba. No travel apps, no Google Flights, no Reddit threads.
Benchmarking the Raspberry Pi AI HAT+ 2 for Local LLMs
Testing the Hailo-10H NPU for home automation: 6-8 tokens/second makes local AI assistants genuinely practical.
Claude Code Browser Automation on Bazzite
A guide to setting up Google Chrome with Claude Code's browser automation on Bazzite, covering both rpm-ostree layering and the recommended distrobox approach.
Hello World
Welcome to my new blog.