I’m looking at DeepSeek’s distilled models for local coding help, but I’m unsure how to choose one for a GPU with 12GB of VRAM. A model that technically loads isn’t necessarily pleasant to use if responses crawl or it struggles with ordinary programming questions.

For people running these locally, which DeepSeek distilled model has felt like a reasonable balance of coding quality and speed? I’d be interested in your runtime, quantization choice, and the kinds of tasks you tested—not just a benchmark score.