Qwen 3.6 powers local Claude Code

This reads less like a product launch and more like proof that local AI coding setups are crossing from hobbyist novelty into genuinely usable infrastructure. The model is only part of the story; the real unlock is the cache behavior and serving stack that keep agent loops fast enough to stay interactive.

• –The key technical dependency is llama.cpp PR 21793, which the author says was needed to make Claude Code work well with the local backend
• –Reported performance is strong for a local setup: 400 t/s prompt processing and 24 t/s generation on Qwen 3.6 35B A3B Q4KM
• –Prompt-prefix caching matters here because agentic coding reuses long instructions and context repeatedly; without it, local workflows feel sluggish fast
• –The hardware footprint is modest enough to be compelling: a 16 GB RTX 2000 Ada in a tiny machine, kept cool with a custom printed fan hanger
• –If these setups keep improving, the practical gap between hosted coding agents and self-hosted ones keeps shrinking for serious developers