Local AI, on purpose
Local models, with the janky left in.
A home-lab notebook: what runs on the machine in front of you, what breaks, and what the IronClaw bench actually measured.
The build log sits underneath. These are the shelves.
One break, one save, from the IronClaw machine.
warn
Featured pitfall
GPT-OSS passed 9 of 16 current-contract tool tasks in one controlled evaluation.
MeasuredWhat it taught meThe tokens-per-second number was not the job. Nine out of sixteen is what that tool contract actually scored.
ok
Featured win
GPT-OSS 20B, MXFP4, llama.cpp, on the RTX 5060 Ti. Generated tokens per second.
Scale, 0 to 120 tok/s
GPT-OSS 20B
MXFP4, llama.cpp
95.72 tok/sMeasured
Qwen3-Coder 30B
quantized Q3 variant
~82 to 84 tok/sMeasured
Gemma 4 26B-A4B
IQ3_XXS
~77 to 78 tok/sMeasured
Qwen3.5 35B-A3B
IQ3 variant
~64 to 68 tok/sMeasured
Rounded observations from different sessions, not a standardized comparison. A range bar uses the midpoint, which is not another measurement.
A GTX 1080 out of the closet, Windows 11 Pro, Ollama, and Llama 3 8B. The first real build log.
Docker, Tailscale, a firewall, and an 8 GB card. The second real build log.
OpenClaw never fully ran. The name that stuck was IronClaw. The third real build log.
A WSL fix wiped the containers. A 16 GB card and a fresh Windows install came after. The fourth real build log.
GPT-OSS 20B on 16 GB, a switch from Claude to GPT, and the lesson that the build was the problem. The fifth real build log.
Dots on the $100 plan, a return to $20, and Grok Bot building this site. The sixth real build log.
A Thursday PHP flaw, a weekend of updates, and a plan to watch security news from the machine. The seventh real build log.
Coming soon
The list isn't open yet. Check back soon.