Qwen-Image-3.0 ships — 4,500-token prompts, but no open weights
Alibaba's text-rendering leap reads type down to 10px and speaks 12 languages — yet ships no weights this time. Plus a new MIT Mac app for local models.
Copy markdown
Qwen-Image-3.0 is live in Qwen Chat
Alibaba's new image model accepts prompts up to 4,500 tokens (versus roughly 1,000 in 2.0), renders legible text down to 10-pixel type, spans 12 languages, and lays out full web pages, infographics, and live weather graphics in a single pass. Try it now in Qwen Chat.
But no weights, no API, no benchmarks
Qwen-Image 1.0 and 2.0 both shipped open weights plus a technical report; 3.0 ships neither — no downloadable model, no API endpoint, and no benchmark numbers. If your pipeline self-hosts Qwen-Image, there's nothing new to pull yet, and the text-rendering claims stay demo-only for now.
Nativ turns your Mac into a local model server
New MIT-licensed macOS app (v0.0.1, Apple silicon) runs MLX models with a chat UI, a model manager, and a local server that speaks both the OpenAI and Anthropic APIs — so you can point Claude Code or Codex at a model running on your own laptop. Free on GitHub.