OpenCode vs Codex: local LLM coding tests

Gary Sims tests whether OpenCode can replace Claude Code or Codex when paired with a local LLM. The task is to build an interpreter for New Scrippy, his typeless language with functions, assignments and conditional statements. He reports that Codex with GPT-5.5 completed the task in six minutes and passed his example scripts.

For OpenCode, Sims uses Ollama's launch feature, which handles configuration after OpenCode is installed. The setups include a Qwen 3.6 35B model on Jetson Thor and a 27B model on an RTX 5090. He reports speeds of 36, 13 and seven tokens per second across three configurations. The transcript identifies the Q8 model inconsistently, calling it 35B during setup and 27B during the results.

All three OpenCode runs reportedly produced interpreters that ran the README example. A subsequent ChatGPT code review flagged memory leaks and other defects. Sims says the 27B Q4 output had critical memory bugs and treated division as multiplication. His local runs took hours, sometimes overnight, and required further prompts.

Sims found the Q8 output better than Q4, with more clarification questions and fewer critical bugs. A separate Qwen 3.5 9B Q8 attempt failed. His conclusion is limited to these tests: the coding assistant produced working examples, but its speed and unresolved defects made the results unsuitable for production.