Local models get serious as cloud AI gets restrictive

Someone got Qwen 3.8 27B running on a laptop and used it to interpret model weights (50355251). In the Ask HN on replacing Claude or GPT with a local model (48542100), one commenter says Qwen 3.6 27B dense lands around Claude Haiku 4.5 and sometimes Sonnet. Others run DeepSeek V4 Flash on two RTX Pro 6000 cards at 160 tok/s, or complain that Gemma 4 on an M4 is too slow. Meanwhile Anthropic stopped letting Claude Code subscriptions power OpenClaw (47633396), offering a one-time credit, and users read it as a squeeze on capacity they already paid for.

The through-line is dependence. Rate limits, tool bans and training-on-your-chats are all push factors. Open weights at 27B that match last year's small cloud models are the pull factor.

Nvidia complicates the picture. A report says encrypted firmware on the RTX Spark slows Linux inference to 35% (50362905). The hardware layer is where local AI could get throttled.