Qwen, Gemma vie for M4 coding

The useful takeaway is not "run the biggest model you can squeeze in"; it is that 24GB Apple Silicon is a capable local inference box, but still a compromise for coding agents.

• –Qwen gets the strongest community nod, with 14B-class models fitting comfortably and 30B-ish quantized models possible but tighter on memory and context
• –Gemma is framed as a safer fit on 24GB, while larger Qwen variants may trade too much speed or headroom for quality
• –For actual heavy coding, commenters still favor Claude, Codex, or other paid remote models, using local LLMs for drafts, private tasks, or backend automation
• –MLX support matters on Mac because memory bandwidth and Apple Silicon-native loaders can make the difference between usable and frustrating