Stream

Today, we’re releasing the first version of our Vision AI framework! It's 100% open-source and has over 10 out-of-the-box plugins for @krispHQ, @ultralytics, @cartesia, @OpenAI, @GoogleAIStudio, and more! With Vision Agents, developers can build fast, low-latency vision AI applications running on Stream from day one with just a few lines of Python ⚒️

We’re excited to collaborate with @TencentRTC to support their WebRTC transport in Vision Agents! If you’re building agents in China and across Asia, join our early access program to try the Tencent RTC transport: visionagents.ai/integrations/e…

Learn more 👇. getstream.io/blog/best-on-d…

Discover, compare, and use VibeVoice, Qwen3-TTS, Neu TTS, TADA TTS, Pocket TTS, and Kitten TTS for voice generation in your next app.

New text-to-speech (TTS) models are released each week, but it is difficult to know which ones can run privately and locally on edge devices. We have tested and shortlisted the Top 6 On-Device TTS Models for Any AI Project below.

What's new: - Anam avatar integration: your agent gets a synced face with interruption handling - LocalEdge: prototype with your mic, speakers, and camera. No Stream account needed - Helm charts for Kubernetes deployment - Deepgram TTS moved to persistent WebSocket - major latency drop - Fixed resource leaks in long-running deployments 100% open source: getstream.io/blog/vision-ag…

Building multimodal AI agents that actually run in production is a different game from demo-ware. Memory leaks, connection cleanup, deployment infra, latency - the boring stuff matters most. Vision Agents v0.5.0 is focused on exactly that 👇 x.com/visionagents_a…

Full roadmap: getstream.io/blog/moderatio…

What's coming in 2026: - Full appeals workflow in the dashboard - Message audit trails for every edit and action - Standalone moderation SDKs (Node, Go, JS) - Mixed-language detection in a single message - Moderator wellbeing tools (session limits, break enforcement)

Content moderation used to be a keyword blocklist. Now it needs to handle real-time video, code-switching across languages in one message, and sentiment analysis - all before a single user sees it. Details in 🧵

You can use Stream Video via Vision Agents too!

Been experimenting with running local models with @visionagents_ai and it's been the most fun I've had in a while. Here's @googlegemma 4 (MLX) running locally on my Macbook via our @huggingface integration doing some realtime video/frame understanding

Build your own by following our tutorial! getstream.io/blog/restauran…

The telephone has been around since 1876. Even 150 years later, it still has more to give! We built an AI agent that answers calls for a restaurant, checks the menu, finds availability and books a table. Build your own with @twilio, @turbopuffer (RAG) + @visionagents_ai 👇

Get started with live audio + video moderation ↓ getstream.io/moderation/doc…

Stream just shipped live audio + video moderation for calls and livestreams. 🚀 - Video keyframes scanned in real-time for nudity, violence, weapons - Audio transcribed and analyzed via LLM or NLP harm labels - Escalating actions: warn, blur, mute, kick Breakdown in thread.

🔗 Build a TikTok Live Shopping clone: getstream.io/blog/build-tik…

How to build a TikTok-style livestream shopping app — Stack: - Next.js - Stream Video (livestreaming) - Stream Chat (in-stream messaging) - Firebase (auth + products) Core pieces: - Create livestream sessions - Attach a chat channel per stream - Sync users via Firebase - Render video + chat side-by-side - Add product data + in-stream UI Final product: live video + real-time chat + shopping in one experience Full walkthrough in thread ↓

Huh, so that's why text is called a string

Here is Mistral's voxtral-mini-tts-2603 in action: Multilingual voice cloning agent. You can clone any voice in minutes: