Open weights and local models start to look usable
Meta's Muse Spark Open got a warm reception, and the praise was specific. Commenters liked that Meta worked to make it run on every runner, from Unsloth to MLX. One commenter said Zuckerberg's original Llama thesis is looking more correct by the week, because the rent market on intelligence never materialized. Unless you own the single frontier model, you can't charge rent.
The Ask HN thread on replacing Claude or GPT with a local model for daily coding gave real numbers. One person runs Qwen 3.6 27B dense and rates it about equal to Claude Haiku 4.5, maybe Sonnet on some tasks. Another runs DeepSeek V4 Flash on 2x RTX Pro 6000 Blackwell at 160 tok/s. Others said Gemma 4 on an M4 was too slow, and that tooling for picking and running models is still thin.
The pattern here: local is now good enough for a real slice of work, but not yet a full swap. The gap is speed and tooling, not raw quality.
So what?
If you build on frontier APIs, your margin and your risk both depend on one vendor's pricing. Design your product so the model is swappable, and test an open-weight model on your actual workload now, because the quality gap you assumed last year may be gone.