Local models still can't fully replace cloud AI for coding

A thread asking whether anyone has replaced Claude or GPT with a local model for daily coding work got a nuanced set of answers. The honest summary: not really, not yet. People with serious hardware (2x RTX Pro 6000 Blackwell, Apple M4) are getting workable results on specific tasks, but token speed and quality gaps with frontier models remain meaningful. Qwen 3.6 27B dense got the most positive mentions, described as roughly matching Claude Haiku 4.5 on some tasks.

The pattern here is that the local model question is really several different questions bundled together: is it good enough for my specific tasks? Is it fast enough to not break my flow? Does my hardware support it? The answers to all three are improving but not uniformly. People with specific, well-defined coding tasks are closer to local parity than people doing open-ended agentic work.

Anthropics move to cut off third-party harnesses from Claude subscriptions is adding urgency to this question. If you are building on Claude Code or similar tools, and Anthropic can shut off your access or change pricing, local alternatives become a hedge worth taking seriously even if they are not yet full replacements.