Best LLMs for OpenCode - Tested Locally
The author tested several locally hosted large language models (LLMs) on the AI developer tool OpenCode and compared their performance on two tasks: creating a CLI tool for the IndexNow protocol and preparing a website migration map.
Why it matters
This article provides valuable insights into the performance of different LLMs on AI development tasks, which is crucial for developers choosing the right tools for their projects.
Key Points
- 1Qwen 3.5 27b Q3_XXS on llama.cpp is the clear winner for local OpenCode use, delivering a complete working project with all tests passing
- 2Qwen 3.5 35b on llama.cpp is fast for coding but has reliability issues, requiring validation of its output
- 3Bigpicle from OpenCode Zen surprisingly performed well, pausing to search for the IndexNow spec before coding
- 4GPT-OSS 20b requires high thinking mode to be capable, but still fails on structured tasks
Details
The author tested several LLMs on OpenCode, a promising AI developer tool, to see how they perform on two tasks: creating a CLI tool for the IndexNow protocol and preparing a website migration map. The clear winner was Qwen 3.5 27b Q3_XXS on llama.cpp, which delivered a complete working project with all tests passing. Qwen 3.5 35b on llama.cpp was also fast for coding but had reliability issues, requiring validation of its output. Surprisingly, Bigpicle from OpenCode Zen performed well, pausing to search for the IndexNow spec before coding. GPT-OSS 20b required high thinking mode to be capable, but still failed on structured tasks. The author emphasizes that for OpenCode, instruction-following and tool-calling quality matter more than raw speed.
No comments yet
Be the first to comment