Qwen 3.8, GLM and LM Studio: a week of local coding

This seven-day experiment tests whether a local LLM can handle everyday coding on an ASUS NUC 16 Pro Mini PC with 64GB RAM. ASUS supplied the hardware and sponsored the video. The presenter uses Ollama and LM Studio, which he describes as available on Windows, Mac and Linux.

The comparison includes Qwen 3.8 at 27 billion parameters, Muse Glimmer and GLM 4.7 Flash. In short chat tests, the presenter reports about 12, six and 22 tokens per second respectively. These are results from his setup, not general performance guarantees. OpenAI's 120-billion-parameter OSS model did not fit within his available resources.

The setup walkthrough connects LM Studio to VS Code through a custom endpoint. Initial chats fail when tool instructions consume the available context. Disabling tool calls and increasing the context window gets the chat working, but the presenter finds that larger contexts demand more resources. His Claude Code test takes roughly three minutes to answer a greeting.

OpenClaw also responds too slowly for his intended use. He switches to cloud models for orchestration and assigns local models background work through OpenClaw and Hermes. His preferred outcome is a dedicated AI agent workstation with a mix of cloud and local processing. He remains dissatisfied with local models for sustained work on larger codebases.