M4 Max MacBook benchmarked for OpenCode, Qwen3

The M4 Max's unified memory remains the definitive "killer feature" for running 30B+ parameter models at usable speeds on consumer-grade hardware.

• –Benchmarks of ~89 tokens/sec on Qwen3-30B-A3B confirm that MoE models are the sweet spot for high-performance local coding agents.
• –OpenCode is emerging as the premier model-agnostic TUI for developers seeking a local-first alternative to proprietary agents like Claude Code.
• –While 32GB of RAM is viable for 30B models, the community increasingly recommends 64GB+ to accommodate the long-context windows required for multi-file codebases.
• –Switching to MLX-native runners provides a significant 30-50% performance boost over llama.cpp for Qwen models on Apple Silicon.