Major AI providers are ditching subscription plans for usage-based pricing, making those "vibe-coded" side projects suddenly expensive. Anthropic toyed with dropping Claude Code from affordable tiers while Microsoft moved GitHub Copilot to a purely usage-based model. Developers now face real costs for API calls that once came cheap.

The alternative: run models locally. Alibaba's Qwen3.6-27B packs "flagship coding power" into a package small enough for a 32 GB M-series Mac or 24 GB GPU. Recent advances in reasoning capabilities, mixture-of-experts architectures, and improved tool calling mean smaller models can now handle real coding tasks.

Setting this up requires Llama.cpp (or alternatives like LM Studio, Ollama, or MLX). The article recommends specific parameters: temperature=0.6, top_p=0.95, and a context window as large as memory allows. For a 24GB RTX 3090 Ti, a 65,536 context window works well with 8-bit key-value compression.

Three agent frameworks work with local models: Claude Code (configured via ANTHROPIC_BASE_URL), lightweight Pi Coding Agent, and Cline. The trade-off is speed and capability versus cost—free local inference beats metered API calls, assuming you have the hardware.