Tutorial outlines how to build a local coding agent using open-weight models and open-source tools

Sebastian Raschka’s tutorial outlines a production-ready local coding agent stack using open-weight models and open-source tools, positioning it as an alternative to proprietary services such as Claude Code and Codex.

The stack consists of a locally served LLM paired with a coding harness that can read files, make edits, run commands, and verify changes, enabling meaningful coding work within a local environment.

Raschka highlights several motivations for local setups, including predictable costs, reproducibility, offline use, and data privacy, particularly for sensitive workflows like document processing.

The guide focuses on using Qwen3.6 with the Qwen-Coder harness, noting that Qwen models are optimized for this harness and can run alongside other harnesses like Codex or Claude Code on the same machine.

A benchmark from Nvidia’s Polar paper (May 2026) is cited to show Qwen3.5-4B’s strong performance in the Qwen-Code harness, with Raschka assuming similar optimization for the newer Qwen3.6 models.

Hardware requirements for Qwen3.6 35B-A3B are specified: approximately 22 GB download size, 30–40 GB RAM, and smooth operation on a Mac Mini with M4 or a DGX Spark. The model is described as the best in its size class based on recent benchmarks shared by Cohere in June.

The tutorial also mentions alternative models like Cohere’s North Mini Code and Gemma 4, which can be used with the Qwen-Code harness, and provides setup guidance using tools such as Ollama, LM Studio, vLLM, SGLang, and MLX.