Memory, hints, and retrieval help small models reason. Attention limits, robustness, and long context. Local and edge AI get faster, lighter. MCP-powered agent integrations grow up. DIY LLMs, from $100 chats to RL coders