Vidu S2 ships real-time AI avatars and live video editing via API
ShengShu's Sept 16 launch has a public API for 720p interactive characters and live restyling. Plus a 35B model run from SSD, and Spain's first agent breach.
Copy markdown
Real-time 720p avatars you can talk to
Vidu S2-Avatar renders a live digital character at 720p and 25-42 fps, driven by voice, text, and a reference image, up from 540p in S1. That's an interactive presenter, game NPC, or a talking face for your agent, generated on the fly rather than pre-rendered.
Live editing on a stream, not a file
Vidu S2-Editing rewrites an incoming video stream in real time across four tasks: style transfer, outfit swaps, subject replacement, and background replacement, keeping garment edges, textures, and occlusion intact. Think live virtual sets or wardrobe swaps on a webcam feed.
The API and demo are live today
You don't have to wait: there's a playable demo at vidu.com/vidu-stream and documented endpoints at platform.vidu.com/vidu-stream/doc. Pricing isn't posted yet, but real-time avatars or on-the-fly restyling are wireable into an app this weekend.
Edge0 streams a 35B model off your SSD
Edge0-35B-A3B (Apache-2.0, shipped Sept 16) runs a 256-expert MoE on Apple Silicon in ~2.9GB of RAM at ~15-18 tok/s, using a 'prerouter' that predicts the next layer's experts one token ahead so weights load while the current token computes. It trails its fp16 Qwen teacher by 3.9 points across five benchmarks, a real 35B on a laptop with framework and checkpoint open.
Spain logs the first autonomous-agent data breach
Spain's AEPD published its first breach carried out by an AI agent with no human at the wheel: it ran recon, logged in, found and exploited an app flaw, altered personal records, and pulled invoices. The regulator's message to builders: patch faster, tighten identity, and lock down what your agents are allowed to do.
Elsewhere: Mistral now powers Firefox's Smart Window
Mozilla wired Mistral models into Firefox Smart Window (beta) with zero data retention and regional-language tuning, live in France and North America. Consumer-facing, but a notable open-model design win and a marker for where privacy-first, in-browser AI is heading.