Local AI Models Are Getting Genuinely Good

Multiple threads today circled the same observation: local open-weight models are now at a quality level where they're worth using for real development work. Story 48721903 specifically calls out Qwen 3.6 27B as the sweet spot for local development, with a comment already teasing Qwen 3.7 27B. Story 48722052 covers Ornith-1.0, billed as a self-improving open-source model for agentic coding, though commenters immediately poked holes in the claims and suggested it's a fine-tuned Qwen or Gemma. Story 48727116 covers LongCat-2.0, a 1.6T MoE model built on non-Nvidia ASIC clusters.

The pattern: the conversation has shifted from 'can open models compete' to 'which open model is best for which use case.' The benchmark-gaming accusations in the Ornith thread are a tell. When the models get good enough that people feel the need to fake results, the category has arrived.

The dissenting note from the Qwen thread: AI companies are buying up all available silicon, making it increasingly expensive to run frontier-class models at home. The gap between what you can run locally and what the big labs run is closing on quality but widening on compute access.