Local models vs. cloud LLMs: still not there for most
A thread asking whether anyone has replaced Claude or GPT with a local model for daily coding got a genuinely honest set of answers. The consensus: Qwen 3.6 27B dense is the closest to viable, roughly matching Claude Haiku 4.5 on many tasks. But speed on consumer hardware is still a bottleneck, enterprise tooling for model selection is absent, and the reasoning models that would handle hard coding tasks are too slow at 0.7 tokens per second on anything short of an RTX Pro 6000 Blackwell rig.
This connects to the Google DeepMind WeatherNext 3 thread, which highlighted a similar gap: powerful open models exist but adoption in production environments lags because no one has built the harnesses and operational tooling. The 'model is one part of the system' observation in the agent testing thread makes the same point. Having a good model is necessary but far from sufficient.
One commenter offered the most honest local-model alternative: a human brain running on caffeine. The thread didn't treat this as a joke.
So what?
If you're building a product that depends on low-latency local inference for cost or privacy reasons, the hardware requirements today make it a bad default for most teams. Plan around cloud APIs with smart caching and context management rather than betting on local inference catching up in the next 12 months.