LLM Hacker News 15h ago 2 min read
Daniel Vaughan’s Gemma 4 writeup tests whether a local model can function as a real Codex CLI agent, with the answer depending less on benchmark claims than on very specific serving choices. The key lesson is that Apple Silicon required llama.cpp plus `--jinja`, KV-cache quantization, and `web_search = "disabled"`, while a GB10 box worked through Ollama 0.20.5.