Qwen-Image-3.0 nails tiny text and layouts, but ships no weights
Alibaba's chat-only model takes 4.5K-token prompts for one-shot mockups. Plus: Claude Code backgrounds reviews, DeepSeek sunsets old endpoints Jul 24.
Copy markdown
One prompt, a finished multi-panel layout
Qwen-Image-3.0 takes prompts up to 4,500 tokens — 4.5x its predecessor — and returns nine infographic panels, UI mockups, or storyboards as a single stitched-free image, with text legible down to ~10px and LaTeX across 12 languages. It's live in Qwen Chat with API trials open, so you can prototype production visuals today.
The open-weight champ shipped this one closed
Qwen 1.0 and 2.0 shipped Apache-2.0 weights; 3.0 arrives chat-only with no downloadable weights, no benchmarks, and no model card. You can't self-host or fine-tune it, and there's no way to verify the quality claims — a notable reversal for one of open weights' biggest backers.
Claude Code moves /code-review to the background
Claude Code 2.1.218 runs /code-review as a background subagent, so your main session keeps working while it reviews. `claude mcp list` and /mcp now surface the HTTP status when an MCP server fails to connect, making a dead connector faster to diagnose.
DeepSeek's original API endpoints go dark Jul 24
DeepSeek retires its legacy `deepseek-chat` and `deepseek-reasoner` endpoints after July 24, 15:59 UTC — pinned calls start returning 404s unless you move to the current V4 model names. The non-preview V4 stable is expected around the same window.