Qwen vs Claude: coding tests with LM Studio and VS Code

The speaker tests a local LLM coding setup in LM Studio and VS Code, then compares the results with Claude Sonnet 5 through API billing. His Windows machine has an RTX 4090 with 24 GB VRAM and 64 GB RAM. He also reports experiments on an M5 Max with 64 GB unified memory, where acceptable speed still limited his model choices.

Qwen 3 Coder 30B, a mixture-of-experts model with 3 billion active parameters, reaches roughly 75 to 83 tokens per second during the demonstration. It produces a working Tetris game, but leaves JavaScript files unfinished for a platformer. The dense Qwen 3.6 27B runs at about 30 to 32 tokens per second. Its repair attempt encounters a browser-testing error in the coding harness. These are results from the speaker's setup, rather than general performance guarantees.

The setup walkthrough covers the LM Studio server URL, model ID and context size in VS Code. The speaker demonstrates Q8 K/V cache quantization and flash attention to reduce memory use, and warns that excessive context or other GPU workloads can cause severe slowdowns. He also reports feature limitations when connecting local models to preferred editors such as Cursor.

In this comparison, Sonnet 5 completes both games on the first attempt. The speaker prefers cloud models for daily coding, while keeping local AI available for privacy or backup use. A sponsored segment introduces MindsHub Cowork and its model router.