Usage-based pricing killing your vibe - here's how to roll your own local AI coding agents

The Register · 2 months ago

Developers can run local AI coding models like Alibaba's Qwen3.6-27B on their own hardware to avoid rising costs from cloud-based coding assistants like GitHub Copilot and Claude Code. Qwen3.6-27B requires 24 GB GPU VRAM or 32 GB unified memory on M-series Macs, along with specific hyperparameters (temperature=0.6, top_p=0.95) and can be deployed using Llama.cpp with agent frameworks like Claude Code, Pi Coding Agent, or Cline. While smaller models lack the capabilities of frontier models like GPT-4.7 or Claude Opus, they can handle basic coding tasks like building simple web apps and debugging existing code bases when given proper human oversight and sandboxing.