Local and open models are now 'enough' for daily coding

In the Ask HN about replacing Claude or GPT with a local model, people gave real setups. One runs 2x RTX Pro 6000 Blackwell with DeepSeek V4 Flash at 160 tok/s raw. Another says Qwen 3.6 27B dense is about Haiku 4.5 level, maybe Sonnet on some tasks. Someone is waiting on antirez's ds4 to stabilize on Strix Halo. Gemma 4 on an M4 was too slow to bother with. In the MiniMax thread, commenters said DeepSeek and the GLM 5.3 series 'feel like enough' for a lot of tasks.

An earlier thread adds the other side of the push. Anthropic stopped letting Claude Code subscriptions power OpenClaw, citing 'outsized strain on our systems.' Commenters read that as 'you used the capacity you paid for.' One said they hit rate limits within 1:45 on afternoons.

The through-line: closed-vendor limits and pricing pushes people to try alternatives, and open models have reached the point where trying them isn't a punishment. Hardware is still the barrier, and the best setups cost serious money.