Local AI Coding Agent: Qwen3.6 with Qwen-Code vs Subscriptions to Codex and Claude Code
OpenAI
Anthropic
Cohere
Meta
Sebastian Raschka's tutorial explains how to set up a local coding agent using open-weight LLMs like Qwen3.6 and harnesses like Qwen-Code, Codex, or Claude Code. It highlights benefits such as cost, privacy, reproducibility, and offline use. The article compares performance, memory usage, and speed of Qwen3.6 35B-A3B and Cohere's North Mini Code.
Sebastian Raschka provides a tutorial on setting up a production-ready local coding agent using open-weight LLMs and local coding harnesses. He uses Qwen3.6 with Qwen-Code as the primary combination, noting that Qwen models are optimized for Qwen-Code. Alternative harnesses include Codex, Claude Code, and Cline. He recommends using Ollama to serve models like Qwen3.6 35B-A3B (22 GB download, 30-40 GB RAM) or the similar-sized North Mini Code 1.0. A speed benchmark shows Qwen3.6 and North Mini generate about 40 tok/sec on a Mac Mini and 30 tok/sec on a DGX with up to 30 GB RAM at 50k context. He also mentions cloud options for larger models like GLM 5.2. The local setup is transparent, inspectable, free (besides hardware), and provides privacy. He suggests anything faster than 20-30 tok/sec is reasonable for local agent work.
- Сокращения
- RL = Reinforcement Learning — обучение с подкреплением
- RAM = Random Access Memory — оперативная память
- GPU = Graphics Processing Unit — графический процессор
- GUI = Graphical User Interface — графический интерфейс пользователя
Source: Sebastian Raschka —
original