Local LLMs Still Not There for Daily Coding
A thread asking whether people have replaced Claude or GPT with a local model for daily coding produced an honest mixed picture. Some builders are running Qwen 3.6 27B dense and calling it roughly equivalent to Claude Haiku 4.5. Others tried Gemma 4 on Apple M4 and found tokens per second too slow to be practical. The person with two RTX Pro 6000 Blackwell GPUs running DeepSeek V4 Flash at 160 tokens per second is not a typical setup.
The pattern is that local models work for specific tasks and specific hardware, but the gap with frontier cloud models in responsiveness and capability is still real enough that most working developers have not made the switch. The enterprise tooling to pick and run appropriate local models is also called out as missing.
One commenter's response was to suggest 'a local model on a natural wetware substrate' with caffeine and a quiet environment, which landed well because it captures how the tradeoffs actually feel to people doing real work.
So what?
If you are building developer tools or coding assistants, do not assume local model parity with cloud models is close enough to matter for your product today. The people who have made the switch successfully have unusual hardware. For most users, cloud models are still the practical choice, which means API dependency and cost remain real constraints to design around.