I ran Qwen3.5 locally instead of Claude Code. Here’s what happened.

A hands-on experiment running Qwen3.5 models locally on consumer hardware as an alternative to Claude Code. Using LM Studio and the Continue VS Code extension, three Qwen3.5 variants (9B q5_1, 9B distilled, and 4B) were tested for code analysis on a small Python project. The models provided useful high-level suggestions and code snippets, but consistently failed when attempting to autonomously apply changes to the codebase — crashing, looping, or even attempting to delete project files. The conclusion: local LLMs work reasonably well for advisory feedback on code, but are not yet reliable for agentic tool use requiring large context windows.