Local AI models: still not good enough, but getting closer

A thread asking whether anyone has replaced Claude or GPT with a local model for daily coding drew sharp, experience-based answers. The consensus: not yet, with caveats. Users running Qwen 3.6 27B dense reported results roughly comparable to Claude Haiku 4.5. Someone with 2x RTX Pro 6000 Blackwell cards running DeepSeek V4 Flash got 160 tokens per second, which they described as usable for their workflow. Apple M4 users found token speed significantly lower than cloud offerings. The enterprise tooling gap was flagged repeatedly: knowing which model to use and how to run it efficiently remains unsolved.

Separately, a solo dev posted achieving 44% on ARC-AGI-1 for just 67 cents of compute. The comments immediately questioned the benchmark: multiple people noted that LLM score improvements are now driven mainly by post-training on synthetic data, meaning models are learning to solve ARC tasks specifically rather than demonstrating general reasoning. 'Benchmaxxing' was the term used.

The through-line: local models are becoming viable for specific, narrow tasks, but the gap in tooling, speed, and general capability keeps cloud APIs dominant for most serious coding work. And the benchmarks that were supposed to tell us when AI 'gets there' are themselves being gamed, leaving the community without a reliable signal.