Qwen2.5-Coder 14B stays local sweet spot

Qwen2.5-Coder 14B is the boring answer, which is usually the right one here: it is big enough to write decent React and API glue, but small enough to stay usable inside Ollama and Continue on 24GB RAM.

• –Qwen2.5-Coder was built for code assistants and code repair, and the official 14B variant keeps the model in the range where local iteration still feels interactive.
• –Qwen3-class and DeepSeek Coder models are stronger at reasoning and harder tasks, but they are less convincing as an everyday local helper when latency and output discipline matter more than benchmark bragging rights.
• –For piece-by-piece full-stack work, the best local model is usually the one that gives clean, direct completions instead of clever but bloated architecture.
• –If the choice is simple: 7B for speed, 14B for the default sweet spot, and bigger only if you can tolerate slower, more quantization-sensitive runs.