OpenCode users name top local models

Running massive models like GPT-OSS 120B on CPU for "overnight coding" is a viable strategy for complex planning, but backend configuration is critical for reliable tool calling.

• –GPT-OSS 120B is a community favorite for planning but often requires specific chat templates (e.g., llama/jinja) to prevent tool-call failures.
• –Qwen 3.5/3.6 models are noted for higher reliability in autonomous "act" phases compared to larger, less refined models.
• –Backend choice matters: Users report that switching from Ollama to llama.cpp or vLLM often resolves JSON formatting issues in tool execution.
• –Memory management is the primary bottleneck; 256GB RAM allows for 256k+ context, which is essential for OpenCode's project-wide analysis.