Qwen-Image-2.1-Turbo: 8-step open image gen with transparent PNGs
Qwen ships a fast open image checkpoint, Google's EmbeddingGemma 2 tops HF trending, and Claude's Managed Agents learn to orchestrate each other.
Copy markdown
8 steps instead of 40, with transparent output
Qwen-Image-2.1-Turbo is an official fast checkpoint of the ~7B Qwen-Image-2.1 that renders in 8 denoising steps instead of 40 — a ~5x cut in step count — at up to 2048px, and emits transparent RGBA PNGs directly so you skip the background-removal pass. It runs locally through Diffusers or ComfyUI (~32GB weights) and does multi-reference composition. The catch: it's under the Qwen Research License — fine to evaluate, not to ship in a commercial product.
Google's open embeddings grow eyes and ears
EmbeddingGemma 2 (Apache 2.0) maps text, code, images, audio, and video into one 768-dim space, with Matryoshka truncation down to 128d, an 8K context, and 100+ languages. It's 740M total but runs text-only at 270M — small enough for on-device RAG — and it's sitting at the top of Hugging Face's trending list right now, scoring 68.46 on MTEB English. Build a fully offline multimodal search box over your own files, no API bill.
Claude's Managed Agents can now run other agents
Anthropic's Managed Agents added Dynamic Workflows in beta: one agent can write a program that spins up many agents in phases — say, reviewing hundreds of documents — and streams progress on workflow_run.* events. You turn it on by adding "workflows": {"type": "enabled"} to the multiagent config. It's a cleaner path than hand-rolling your own fan-out orchestration on top of the API.