BREAKING: Apple and Meta Just Teamed Up. Nobody Saw This Coming

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: BREAKING: Apple and Meta Just Teamed Up. Nobody Saw This Coming · Channel: Hyperautomation Labs · 09:16 · Watch on YouTube

Topics: AI agents, Model inference and optimization, Markets and economy, Local and open-source models

Summary

Meta released Muse Spark Open, an open-weights model with 816 billion parameters that can run locally across multiple Mac Studios. The video describes the model’s agent and multimodal capabilities, its benchmark strengths and weaknesses, and how Apple and Meta worked together to run it using linked Macs. The setup is expensive, but viewers can also try the model through hosted services or use smaller Muse models. The presenter frames the release as a notable return of a powerful American open model and a surprising collaboration between Apple and Meta.

Key points

Muse Spark Open is an 816-billion-parameter model, with about 42 billion parameters active at a time, and its weights are available on Hugging Face.
The model can process text, images, video, and audio, and supports a context of up to one million tokens.
Meta designed it especially for agentic tasks, where AI uses tools and works through steps over time.
According to Meta’s benchmark comparisons, it leads the listed open models on tool use, real job tasks, Humanity’s Last Exam, and visual puzzles, but trails competitors on coding and some overall intelligence comparisons.
The video highlights a jailbreak-resistance score of 0.5 for Muse Spark Open, with lower scores presented as better.
Meta demonstrated the model coding an iPhone app and editing video in DaVinci Resolve without cloud processing.
Apple and Meta linked two Mac Studios with Thunderbolt 5, splitting the model across them so both machines work together.
The local setup is costly: the video cites a starting price of $10,799 for a Mac Studio with M5 Ultra and 256 GB of memory, and says four would cost over $40,000.
Alternatives include renting access through services such as OpenRouter, Together, and Fireworks, or running the smaller Muse Glimmer model on a regular computer.
The video says the license is free for almost everyone, with a requirement to label apps as built with Muse Spark Open and a specific condition for API businesses making over $20 million a year.

Chapters

Start	Chapter	Summary
00:00	Apple and Meta’s unexpected collaboration	The video introduces the local Mac Studio demonstration and explains the companies’ history of conflict and Meta’s shift toward closed models.
03:05	What Muse Spark Open can do	The presenter outlines the model’s architecture, context length, agent focus, benchmark results, and demonstrated coding and video-editing tasks.
05:39	How linked Macs run the model	The video explains the shared-memory Mac setup, its reported performance, and the high cost of the hardware.
07:13	Ways to try it and what the release means	The presenter describes hosted access, the smaller Muse Glimmer model, and licensing, then reflects on the significance of the release.

Quotes

Not in a data center. On a desk. — 00:00

Muse Spark Open weights releasing today, runs great on new Mac Studios. — 02:03

That’s how a model with 816 billion parameters only uses about 42 billion at a time. — 03:05

In Meta’s words, two Mac Studios behave like one supercomputer with a terabyte of unified memory. — 06:10

Two companies that couldn’t stand each other found one thing they both want. — 08:15

Concepts and entities

Muse Spark Open, Muse Wonder, Muse Glimmer, Meta, Apple, Mac Studio, Thunderbolt 5, Hugging Face, OpenRouter, DaVinci Resolve, Alexander Wang, John Ternus, Daniel Gross

Full transcript (original language)

00:00 Apple and Meta two companies that have been at war for 5 years just put an 816 billion parameter AI on a desk. Not in a data center. On a desk. This is the photo Mark Zuckerberg posted today. Two Mac Studios, one screen, and a little blue Meta logo sitting right between them. Nearly a million people have already looked at it. Here’s why that matters to you.

00:30 Right now, every time you ask an AI something hard, your words leave your house. They travel to a data center you’ll never see, and you pay rent on every single question. Those two Macs are running what Meta calls the most capable open model an American company has ever released, and nothing leaves the desk. By the end of this video, you’ll know what Meta just released, why Apple of all companies helped, how two Macs can act like one computer, and the cheapest

01:00 way to use it tonight. But first, you need to know why this handshake almost never happened. Go back to 2021. Apple ships an iPhone update with one little pop-up. Ask app not to track. Millions of people tapped it, and that one button hit the engine of Meta’s business, which was ads that follow you around the internet. Meta said the change would cost it about 10 billion dollars in a single year. After that, the two companies barely spoken in public. No joint launches, no

01:32 shared stage. Just years of cold war between two of the most valuable companies on Earth. And there’s a second twist. For years, Meta was the hero of free AI. Its Llama models were the ones anyone could download, run, and change. Then in April this year, Meta launched a brand new family called Muse, and the big models were closed. You could rent them through Meta servers, but you couldn’t download them. Meta did share one small open model in August called Glimmer.

02:03 But 3 days ago, Meta launched Muse Wonder, its strongest model ever. And that one was locked, too. A lot of people decided Meta was done giving its best AI away. Then today, Zuckerberg posts this. Muse Spark Open weights releasing today, runs great on new Mac Studios. Thanks to work with John Ternus and Apple. John Ternus is Apple’s CEO. And 44 minutes later, he replied. Congrats to the team. Amazing to run a model of Muse Spark Open’s caliber and

02:34 scale this efficiently right from your desk. Mac Studio keeps pushing the boundaries of on-device AI. Meta’s chief AI officer, Alexander Wang, told Axios the two teams had been working together quietly for months. He even credited a former Apple executive, Daniel Gross, with helping bring them together. And his summary of the whole thing? Excited for Mommy and Daddy to be playing nice again. Neither company would say if this is the start of something bigger, but the weights are real.

03:05 They’re on Hugging Face right now. So, what is it? Picture a stadium with 256 experts in the seats. Every time the model reads a word, it wakes up only the eight experts who know that word best. Everyone else stays asleep. That’s how a model with 816 billion parameters only uses about 42 billion at a time. The knowledge of a giant with the speed of something much smaller. It reads text, images, video, and even audio.

03:36 It can hold a million tokens in its head at once. That’s about seven long novels or an entire code base. And Meta built it for one job above all. Agents. AI that doesn’t just answer you, but does the work using tools step by step for hours and fixing its own mistakes along the way. But, is it actually good? Here’s Meta’s own scoreboard against the best open models in the world. Kimi K3 and GLM 5.3 from China, Qwen, Mistral, Nvidia’s Nemotron, and more.

04:08 First, the wins. On using real tools, it beats every open model on this chart. On real job tasks, it’s number one. On the brutally hard test called Humanity’s Last Exam, with no internet help, it’s number one again. And on tricky visual puzzles, it’s way out in front. Now, the part Meta didn’t put in the headline. On coding, it loses. In the terminal, Kimi K3 beats it by almost nine points. On real software bugs, it’s behind Kimi and GLM.

04:39 And on the overall intelligence index, two Chinese models still sit above it. So, most capable American open model is a careful centers. It’s the best one made in America. It isn’t the best open model in the world. But, here’s the number that surprised me. On a test of how easily a model can be tricked into doing harmful things, the other models scored between 63 and 86. Lower is better. Muse Spark Open scored 0.5.

05:09 It’s by far the hardest of these models to jailbreak. So, what can it actually do on the desk? Meta gave it two real jobs with no cloud at all. Job one, Ice Cubes, a real iPhone app on GitHub with about 7,000 stars. Running inside Meta’s coding tool, the model fixed bugs, added new features, and built a whole app start to finish on the two Macs. Job two, video editing. It drove DaVinci Resolve by itself.

05:39 It reviewed the footage frame by frame, found the strongest moments, and cropped them into vertical clips, which, honestly, is a big part of my job. Now, here’s why Apple matters. A normal PC keeps the AI on a graphics card, and even the big gaming cards only hold a few dozen gigabytes. Apple’s new M5 Ultra chip shares one giant pool of memory, up to 512 gigabytes in a single Mac Studio. Even squeezed down, this model needs

06:10 about 500 gigabytes. Too big for one Mac. So, Apple and Meta did something clever. They linked two Macs with a Thunderbolt 5 cable, cut every layer of the model in half, and gave one half to each Mac. Both halves work at the same time, then swap answers over the cable. In Meta’s words, two Mac Studios behave like one supercomputer with a terabyte of unified memory. Apple’s own engineers measured it. On two Macs, it writes about 25 tokens a second. With four Macs, almost 37.

06:43 That’s faster than you can read. Now, the bad news. A Mac Studio with the M5 Ultra and 256 gigabytes starts at $10,799. And you’d need four of them. That’s over $40,000. The 512 gigabyte version, which would let you get away with two, doesn’t even have a price yet. It’s expected late this month. So, no, this is not running on your MacBook. This desk is for companies already

07:13 spending serious money on AI every month, who would rather buy the machine once than pay rent forever. If this video is useful, hit subscribe because it’s free. And if you want to back the work, hit join here or the supporter button on Facebook. Now, the good news. You don’t need a single Mac to try it. Door number one. Rent it. It’s already live on OpenRouter, Together, Fireworks, Grok, and more at about a dollar 25 per million tokens in. That’s the same price as Meta’s closed Spark model.

07:44 Door number two, go small. Meta’s little open sibling, Muse Glimmer, fits on a normal computer with about 18 GB of memory, and it was downloaded about a million times in the last month alone. Door number three, build with it. The license is free for almost everyone. You just show built with Muse Spark open somewhere in your app. You only need to call Meta if you sell it as an API and make over $20 million a year doing

08:15 it. So, here’s the big picture. For the last year, the best open AI models came from China. Today, an American company put a serious one on the board. And Apple turned the Mac into the place you run it. Two companies that couldn’t stand each other found one thing they both want. AI that lives on your desk instead of someone else’s server. Now, I want your vote. If you had $40,000, would you buy the Macs and own your AI

08:45 forever or rent it for $1.25 and keep the cash? A or B? Tell me in the comments. Real talk, 90% of you watching haven’t subscribed. On Facebook, subscribe is the supporter button. That is what keeps the lights on here. On YouTube, it’s free and it tells the algorithm to send you more. Subscribe to support us. I put every number from this video on one free page, the Muse Spark open field card,

09:16 the scoreboard decoded, the Mac setup, the price table, and three ways to try it. Comment terabyte and I’ll send it to you. If you’re new to Claude, my beginner guides are in the description. Follow along on Instagram and Facebook for bite-size AI. And the deep dives live here on YouTube. If Apple and Meta are friends now, the next few months are going to be wild.