Kyu Park

Notes and experiments on local AI, agents, and model benchmarks.

LEGO city: one prompt, 3 models →
Gemini 4 Argon
Battle royale game: one prompt, 9 models →
Claude Opus 5.5
Drone flight: one prompt, 11 models →
Claude Opus 5.5
Wizard's Chess: one prompt, 3 models →
GLM 5.3 Flash

Posts

2026-10-22

Two small tools from my Hermes setup, now open source

session-vault archives every AI agent session into one searchable SQLite file; search-fusion is a budget-aware web search plugin for Hermes. Both MIT.

2026-10-22

Opus plans, DeepSeek builds: a cheaper loop for game visuals

Claude Opus reviews screenshots and writes the plan; DeepSeek's dsh agent writes the code. Same improvement for about a third of the cost.

2026-05-11

Running Qwen3.6 27B dense locally on M4 Max: which path works?

Benchmarks of every practical local path for Qwen3.6 27B dense: oMLX, OptiQ, Ollama NVFP4/MXFP4, KV quantization, and long-context behavior. Low vs High Power.

2026-05-10

Local LLM inference on an M4 Max 128GB

Benchmarking local LLM runtimes in Low vs High Power mode. oMLX, Ollama, ds4, Gemma 4, and DeepSeek V4 Flash compared.