Overview
Qwen2.5-Coder is a family of six dense code models trained on about 5.5 trillion tokens of code and code-related text. The 32B-Instruct model was the first open model of its size to match GPT-4o on several code generation benchmarks, and the smaller sizes became the default choice for local autocomplete in tools like Continue and Tabby.
Strengths
- Wide size range, so you can pick a model that fits your hardware
- Fill-in-the-middle support for inline autocomplete
- 128K context on the 7B and larger models
- Strong code repair and code reasoning for its size
Run it locally
# Chat / refactoring
ollama run qwen2.5-coder:14b
# Fast autocomplete
ollama run qwen2.5-coder:1.5bGood fit for
Local IDE autocomplete, offline coding assistants and fine-tuning experiments on modest hardware.
Frequently asked questions
Which Qwen2.5-Coder size should I use?
7B is a good default for autocomplete on a laptop. 14B and 32B give noticeably better results for chat and refactoring if you have 16 to 24 GB or more of GPU memory.
Does Qwen2.5-Coder support fill-in-the-middle?
Yes. The base models support fill-in-the-middle (FIM), which is what editor autocomplete plugins use.