Mistral open-sources Shieldstral, a 3B moderation model you self-host
Write your moderation policy in plain English and run it on one 16GB GPU. Plus ByteDance's real-time watch-listen-speak model and Claude Opus 4.1's retirement.
Copy markdown
One policy prompt, no retraining
Shieldstral treats moderation as a yes/no question you write in plain language at inference time, so you change what gets blocked by editing a prompt instead of fine-tuning. It returns a calibrated probability from a single forward pass, letting you rank or threshold rather than accept a hard label.
A 3B that matches 20B guards
It scores 84.9% average F1 on text-safety benchmarks, level with the 7x-larger GPT-OSS-Safeguard-20B, and 83.8% on multimodal, ahead of OmniGuard-7B's 77.6%. Built on Ministral-3-3B with a Pixtral vision encoder, it screens text and images together.
Apache 2.0 weights, one 16GB GPU
The weights are on Hugging Face at mistralai/Shieldstral-1.0-3B and fit a single 16GB card, so you can put a self-hosted moderation layer in front of any app and stop paying per call to a hosted moderation API.
ByteDance SeedRealtime watches and talks at once
ByteDance shipped SeedRealtime, a full-duplex model that watches, listens, and speaks across continuous audio-video streams at the same time, the missing piece for live voice-and-vision agents that don't wait for you to finish a sentence.
Elsewhere: Opus 4.1 retires, sessions get shared memory
Anthropic deprecated claude-opus-4-1-20250805 on Aug 5 ahead of retirement, so migrate any hardcoded calls to Opus 5. Separately, MemoryPlugin launched a Mac app that syncs Claude Code, Codex, and Cursor sessions into one memory searchable from 21+ tools.