- The Summary AI
- Posts
- 🔥 Meta Releases Voice-First Assistant
🔥 Meta Releases Voice-First Assistant
PLUS: Qwen AI Runs Fast on MacBooks

Welcome back!
Meta unveiled its standalone personal AI assistant app powered by Llama 4, designed to personalize conversations based on user data. With real-time voice, memory, image generation, and integration with Meta smart glasses, it’s part assistant, part social hub. Let’s unpack...
Today’s Summary:
🦙 Meta launches personal AI app
🗣️ NotebookLM expands to 75 languages
🧠 Alibaba releases open-source Qwen3
🔌 Claude adds plug & play integrations
💡 Microsoft unveils Phi-4 reasoning models
✏️ Gemini app gets image editing
🧐 OpenAI fixes ChatGPT flattery bug
🛠️ 2 new tools

TOP STORY
Meta launches personal AI App built on Llama 4
The Summary: Meta has released its standalone AI assistant app, built on Llama 4 and deeply personalized using data from Meta platforms. The app brings voice-first interaction, memory, and image generation to mobile and web. It also doubles as a social space, with a Discover feed sharing how people are using it.
Key details:
Meta AI is now available in a dedicated app
The app uses real time full duplex voice interaction and integrates with Ray-Ban Meta smart glasses
A premium tier with ads is in the works as Meta eyes monetization
Meta will invest up to $72B in AI this year
Why it matters: Meta is turning its AI into a system that works across phones, glasses, and the web, using personal history from its platforms to compete with ChatGPT and Gemini. It’s aiming to become the default interface for digital life, responding faster since it already knows the context. The goal is simple: reduce friction and increase reliance.

NotebookLM now speaks in 75 languages
The Summary: Google’s NotebookLM research assistant is going mobile and multilingual. Its hit feature, Audio Overviews, now speaks in 75 languages, powered by Gemini 2.5 Pro. The new iOS and Android apps debut on May 20, letting users turn research into podcast-like experiences.
Key details:
Audio Overviews now support 75 languages
The new mobile apps for iOS and Android drop May 20, the first day of Google I/O 2025, with offline playback and background listening
Audio is generated using Gemini 2.5 Pro with “metaprompting,” producing synthetic conversations that mimic real podcast hosts
Teachers and homeschoolers already use NotebookLM to convert mixed-language sources into multilingual audio lessons for students
Why it matters: NotebookLM Audio Overview went viral 6 months ago by introducing the idea of AI podcasts. Now, by expanding to 75 languages and going mobile, Google is doubling down on converting dense source material into audio podcasts. From PDFs to lectures, it’s a useful tool for studying, prepping, and absorb complex content.

FROM OUR PARTNERS
Boost productivity with smarter AI prompts
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

OPEN SOURCE
Qwen3 lands with hybrid reasoning and impressive performance on Macbooks
The Summary: Alibaba has launched Qwen3, a family of open-source language models packed with hybrid reasoning, multilingual fluency, and a range of sizes from 0.6B to 235B parameters. The flagship models rival top-tier systems like DeepSeek-R1 and Gemini 2.5 Pro, while the smaller models are proving surprisingly capable. Released entirely under the Apache 2.0 free license.
Key details:
First open model that feels fully usable at speed and quality on a MacBook with 32GB+ RAM, using LM Studio with Qwen3-30b-A3B-MLX-4bit
Qwen3-30B-A3B reaches 70 tokens/sec on a Mac M3 Max; community reports show real-time usability even on common GPUs like the RTX 3060
Support 119 languages including obscure regional tongues
Ranks 3rd globally on the RAG hallucination leaderboard, outperforming OpenAI models in factual consistency
Why it matters: Qwen3 proves that smarter training drives performance. With its thinking/non-thinking toggle, it lets users allocate reasoning compute and adapt to real workflows. More importantly, it delivers fast, high quality performance on consumer hardware, with users reporting 70 tokens/sec speed on a MacBook.

QUICK NEWS
Quick news
Microsoft releases Phi-4 open-source reasoning models
OpenAI explains why ChatGPT became too flattering
Upload and edit your pictures in the Gemini app

TOOLS
🥇 New tools
LLMRefs - Increase brand visibility in AI Search
Midjourney Omni Reference - Precise element control in v7 images
Ztalk - Realtime voice translation for video calls

That’s all for today!
If you liked the newsletter, share it with your friends and colleagues by sending them this link: https://thesummary.ai/