Local coding autocomplete still has a model-size problem
A long-running local setup uses Qwen2.5 7B in a quantized form with the llama.vscode extension for coding . It still works, but it now feels weaker than cloud tools such as Cursor. Qwen3 Coder and Qwen3 Coder Next work, but they are too large for this setup.
The stronger 3090 s are already used for in chat and agent work, leaving only one 3060 card or a local MacBook for . Qwen3 did not work for this use case. Qwen 3.5 and Qwen 3.5 Base technically worked, but they were much slower and worse than Qwen2.5, mostly handling only simple .
Granite 4 also worked but performed worse than Qwen2.5. The practical question is whether there is a better local model for coding besides Qwen2.5 or the larger Qwen3 Coder family.
Key points
- Qwen2.5 7B still works for local coding , but it feels behind cloud tools like Cursor.
- Qwen3 Coder models work, but they are too large for the available hardware.
- The main s are already used for chat and agent workloads.
- Qwen3, Qwen 3.5, and Granite 4 were not better replacements in this setup.
- Local can save money, but only if the small model is fast and good enough.