Open-weight models keep closing the gap

Inkling 2 Small landed with commenters calling it possibly the first US open-weight model good enough for real knowledge work, with one framing it as an 'open weight US based Opus 5.5 within' reach. A Show HN trained a decent text-to-image model for about $20K and released Apache 2.0 weights, and people immediately asked about quantizing it for WebGPU in the browser. Google shipped Gemini Omni 1.2 Flash the same day, and a separate Ask HN on replacing Claude or GPT with local models for coding drew real reports: Qwen 3.6 27B dense judged roughly Haiku 4.5 level, DeepSeek V4 Flash on two RTX Pro 6000s at 160 tok/s.

The through-line is cost collapse. A usable image model for $20K and a coding model that runs on one workstation change who can afford to compete. Complaints are specific and honest: smoothing, mangled limbs, slow tokens per second on an M4. Nobody claims parity, but 'good enough for most work' is now the bar people use.

Anthropic's earlier move to stop Claude Code subscriptions from working with OpenClaw, with commenters calling it a money grab, is part of why local options feel more attractive.