A Thoughtworks technologist shares hands-on experience running local LLMs for agentic coding over four weeks on Apple Silicon hardware (M3 Max 48GB and M5 Pro 64GB). The post systematically covers the key factors affecting viability: RAM constraints, processing power, memory bandwidth, parameter count, reasoning capabilities, tool calling reliability, model format (GGUF vs MLX), quantization levels, MoE vs dense architecture, context window size, runtime choice (LM Studio), and harness compatibility (OpenCode, Pi, Claude Code). Key findings include: Qwen3.6 35B MoE offers the best balance of capability and RAM footprint; reasoning mode can actually hurt performance for smaller models; tool calling remains unreliable but models often self-correct; and the overall experience is still too messy for plug-and-play developer use. A follow-up memo covering specific task results is promised.
