Apple and Meta announced Thursday that Meta’s new open‑weight model, Muse Spark Open, can run on Apple’s Mac Studio machines equipped with the M5 Ultra chip. The collaboration, the first public partnership between the rivals since 2017, leverages Apple’s MLX framework and requires two (or up to four) Mac Studios to host the 816 billion‑parameter model.
Key Takeaways
- Meta’s Muse Spark Open (816 bn parameters, 42 bn active) runs on two Mac Studio units with M5 Ultra chips.
- Each Mac Studio GPU can address up to 512 GB of memory; together the pair hold the full model.
- The setup costs at least $10,799 for a 256 GB‑memory Mac Studio; a 512 GB version is slated for later this month.
- Running locally means Meta does not see user prompts, enhancing privacy.
- This marks the first joint effort between Apple and Meta since 2017 and the first desktop‑class model of its size.
| Metric | Value | Why it matters |
|---|---|---|
| Model size | 816 bn parameters (42 bn active) | Sets a new scale for desktop‑run AI. |
| Hardware required | Two Mac Studio units with M5 Ultra chips | Enables local inference without cloud. |
| Base price | $10,799 (256 GB memory) | Sets a high entry barrier for individuals. |
| Announcement date | Thursday (2024) | Signals renewed cooperation after a seven‑year gap. |
What is Muse Spark Open?
Muse Spark Open is Meta’s latest open‑weight large language model, boasting a total of 816 billion parameters while keeping 42 billion of them active for any given inference pass. Meta positions it as the first model of its class that can be executed on a consumer‑grade desktop, a claim that hinges on the massive parallelism and memory bandwidth of Apple’s M5 Ultra silicon.
How the Model Runs on Mac Studio
The model is split across two Mac Studio machines, each powered by the M5 Ultra chip and the MLX machine‑learning framework. Apple’s GPU architecture can allocate up to 512 GB of VRAM per unit, allowing the two machines together to store the entire weight matrix. Users may link two to four Macs, enabling workloads such as code generation, video analysis, or autonomous agents to run side‑by‑side.
“Muse Spark Open, weights releasing today, runs great on new Mac Studios thanks to work with @johnternus and Apple.” – Mark Zuckerberg on X
Technical Requirements and Cost
To meet the memory demands, each Mac Studio must be configured with at least 256 GB of system memory; a 512 GB configuration is expected later this month. The base configuration with 256 GB of memory starts at $10,799, a price point that reflects the premium hardware needed for high‑end AI workloads. The cost excludes additional Macs that may be required for scaling beyond two units.
Background: Apple‑Meta Relations
The partnership follows a period of tension after Apple’s 2021 privacy overhaul, which Meta estimated shaved $10 bn off its 2022 revenue. According to Meta’s chief AI officer Alexandr Wang, the two companies have been collaborating for several months, with former Apple director Daniel Gross acting as a bridge. The announcement marks the first public collaboration since 2017, though both firms declined to comment on whether this signals a deeper strategic alignment.
Performance and Privacy Implications
Running the model locally eliminates the need to send prompts to Meta’s servers, a point highlighted by Axios. This architecture not only reduces latency but also ensures that user data remains on the device, aligning with Apple’s broader privacy narrative. However, the requirement for high‑end hardware limits accessibility to enterprises or well‑funded creators.
Why it matters
The ability to run a 816‑billion‑parameter model on a desktop blurs the line between cloud‑only AI services and on‑premise inference. For developers, it opens a path to experiment with large‑scale models without exposing proprietary data to external APIs. For the industry, it demonstrates that Apple’s silicon can serve as a viable platform for next‑generation AI workloads, potentially reshaping the hardware market for AI‑centric developers.
Frequently asked questions
What hardware is required to run Muse Spark Open?
At minimum, two Mac Studio units equipped with the M5 Ultra chip and configured with 256 GB of memory each. The GPUs can address up to 512 GB of memory per machine, allowing the full model to be stored across the pair.
Does Meta retain any data from local inference?
No. Because the model runs entirely on the user’s Macs, Meta does not see the prompts or outputs, providing an extra layer of privacy.
How much does a suitable Mac Studio cost?
The base price for a Mac Studio with the M5 Ultra chip and 256 GB of memory is $10,799. A 512 GB memory variant is expected later this month.
Is this the first time Apple and Meta have worked together?
No. The companies have a history of collaboration, but this is the first public partnership announced since 2017.
Can more than two Macs be used?
Yes. The MLX framework allows users to link two to four Mac Studio machines for larger or more concurrent workloads.
Bottom Line
Meta’s Muse Spark Open can now be run on two high‑end Mac Studio desktops, delivering a 816 bn‑parameter model locally while preserving user privacy. The setup demands premium Apple hardware priced from $10,799, positioning the solution for professional developers and enterprises rather than casual users. The collaboration also signals a thaw in Apple‑Meta relations and showcases Apple’s M5 Ultra as a serious contender in the AI hardware arena.
Reporting based on: original source.
Comments
Post a Comment