Smol Coder setup with Ollama and Qwen 3.8

Smol Coder is presented as a free, open source coding assistant designed for models running on your own hardware. The speaker argues that larger coding harnesses consume too much context for smaller models and describes changes to prompts, tools and compaction intended to reduce that overhead.

The tutorial installs Smol Coder with npm and downloads Qwen 3.8 through Ollama. The chapter identifies the model as 27B; the speaker describes an approximately 18 GB download and about 16 GB of VRAM or Mac unified memory. Model choice depends on available hardware. Both the terminal and web interfaces detect models from Ollama and LM Studio. The walkthrough covers reasoning effort, context usage and permission modes. Edit mode asks before running commands, while bypass mode allows file changes and commands.

A browser tic-tac-toe demo introduces the task plan, which the speaker says helps the local LLM resume work after context compaction. For a Minecraft-style game, he first asks ChatGPT for a concise specification, then gives it to the local model. He reports about 20 minutes of uninterrupted work. The result initially has missing textures, which he addresses with a screenshot and another prompt. Image feedback requires a model with vision support. Although local execution can be offline, the demonstrated planning step uses ChatGPT.