SummaryZDNET writer David Gewirtz tested a fully local and free AI coding stack (Qwen 3-coder, Ollama, OpenCode) on a Mac Studio M4 Max with 128 GB of RAM. The system successfully developed a simple Mac utility app but struggled with complex debugging tasks. This led Gewirtz to continue using Claude Code as his primary tool. The 64K token context window emerged as a critical bottleneck in multi-step diagnostic processes. |
Why it mattersDevelopers are increasingly relying on agentic coding assistants. Therefore, knowing where free, on-device models succeed and where they fail is vital for cost and privacy decisions. This hands-on test isolates context length as the decisive constraint. Many marketing claims surrounding local large language models (LLMs) ignore this factor. The test also provides a realistic snapshot for 2026 of the differences between open-weight coding models and closed-source cloud agents. |
Highlights- Qwen 3-coder (80 billion parameters, 3 billion active parameters) successfully handled simple application generation but occasionally leaked Chinese characters into the responses.
- The 64K token context window was the hard limit that caused the failure of the PDF editor debugging task.
- Gemma 4 26 B was used as a multimodal backup for interpreting screenshots, but the model is described as 'excitable' and easily loses focus.
|
Critical takeThe test relies on anecdotes from a single developer working on two projects with a specific hardware configuration. Therefore, the generalization that 'local AI cannot debug' overstates the evidence. Additionally, the 64K token context limit is a configuration choice, not an intrinsic property of the models. The author does not test via comparison whether a larger context window or a different quantization choice would change the outcome. |
This critique was written by the local AI model
Qwen3.8-27B. I chose the story; I do not review each critique before it is published.