Qwen-Image-3.0 lands closed: no open weights, no benchmarks
Alibaba breaks from its open-weights past, leaving self-hosters stuck — as Nativ brings local inference to Macs and Claude Code caps runaway subagents.
Copy markdown
Qwen-Image-3.0 goes closed
Alibaba's third-gen image model reads 4,500-token prompts, renders text as small as 10px legibly, and spans 12 languages — but it ships only inside Qwen Chat, with no downloadable weights, no public API, and not even a benchmark or technical report. Both Qwen-Image 1.0 and 2.0 shipped Apache-2.0 weights, so the flagship just pulled up the ladder for anyone who fine-tunes or self-hosts.
Run frontier models locally on your Mac
Prince Canuma (of MLX-VLM) released Nativ, a free MIT-licensed macOS app that bundles an mlx-vlm server, auto-detects MLX models already in your Hugging Face cache, and exposes both OpenAI- and Anthropic-compatible localhost endpoints. Point Claude Code or any OpenAI SDK client at it and run vision + chat models fully offline on Apple silicon — no cloud, no per-token bill.
Claude Code reins in runaway subagents
Claude Code 2.1.217 caps concurrent subagent spawns at 20 by default (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS) and stops subagents from spawning their own nested subagents. It also fixes --max-budget-usd failing to halt background subagents and patches a memory leak from truncated MCP tool outputs — worth the update if your agents have been fanning out and burning budget unchecked.