Ollama, LM Studio fuel local coding debate

This is really a stack-selection post, not a model review: on 16GB unified memory, the practical ceiling matters more than the brand name on the checkpoint.

• –Qwen2.5-Coder still looks like the safest default for this class of MacBook, especially in GGUF quantized form and with smaller contexts.
• –DeepSeek’s MoE lineup is tempting on paper, but the flagship family is far too large to treat like a normal 14B local model.
• –Continue is still the most credible VS Code path for agentic, multi-file coding; Zed and Void may be improving, but the workflow question is broader than editor parity alone.
• –Throughput on an M1 Pro will be good enough for interactive coding, but the UX will hinge on memory pressure, quantization, and context length more than raw token speed.