Qwen 3.6 with llama.cpp and OpenCode: local coding setup

This tutorial builds a local LLM coding workflow with Qwen 3.6 27B and 35B-A3B, llama.cpp, and OpenCode. llama.cpp serves the models through an OpenAI-compatible API; OpenCode gives the coding assistant access to project files and a terminal. The presenter uses VS Code to test and edit the generated projects.

Two examples compare the models: a Three.js browser car game and a fitness portfolio built with React, Vite, and Tailwind CSS. The car game exposes control and wheel problems in the 35B-A3B result. A separate comparison adds Anthropic's Frontend Design skill. The presenter prefers the resulting layouts, though the 27B model initially misses the skill and needs an explicit mention.

Setup covers the UD-Q4_K_XL MTP variants, a 65,536-token context, automatic GPU offloading, and a local endpoint on port 8080. Building llama.cpp from source is covered in a separate video. The demonstrated machine has an RTX 5060 Ti with 16 GB VRAM and 64 GB system RAM; inference uses both because the models do not fit entirely in VRAM. The presenter reports roughly 10 versus 43 tokens per second after increasing VRAM headroom to about 1 GB. These are results from his setup, rather than general performance guarantees.

The guide finishes with OpenCode configuration, its web interface, VS Code integration, and optional MCP servers. Inference runs locally; optional MCP tools such as web search can involve external services whose hosting and privacy are unspecified.