Explore AI
Writing and reviewing code
Autocomplete, refactoring, explaining unfamiliar code. Punishes small models harder than chat does.
Autocomplete, refactoring, explaining unfamiliar code. Punishes small models harder than chat does.
What we would try first
- Qwen3 Coder 30B A3B 20 GB at Q4 Purpose-built for code, Mixture-of-Experts so only ~3B activates per token. Fast and unusually good at long files. One of the most-downloaded quantised models on the hub.
- Qwen3 32B 24 GB at Q4 The best quality you can run on a single consumer card. Noticeably better at coding and multi-step tasks than any 8B model.
- Qwen3 8B 8 GB at Q4 The default answer for most people with 16 GB or a 12 GB card. Start here, and only look elsewhere if it fails at something specific.