Apple, Meta Pair Up: Muse AI Needs $21,598 in Macs [2026]

Apple and Meta, two companies that have spent a decade trading barbs over privacy and app store rules, confirmed a technical collaboration on October 15, 2026 that lets developers run Meta’s new Muse Spark Open model locally on Mac Studio desktops. The news landed first through CNBC, which reported the partnership on video, and spread quickly across tech outlets covering Meta’s open-weight model push.

Muse Spark Open is an 816 billion parameter model with 42 billion active parameters, a 1 million token context window, and native multimodal support, according to Meta’s own research blog. Meta calls it the most capable US open-weight model to date, and the company says it is built for coding, software engineering, long-horizon agent work, and multi-step reasoning. What makes the October 15 announcement notable is not the model alone. It is the fact that Apple’s hardware and software teams helped make it runnable on Mac Studio computers equipped with Apple’s own M5 Ultra chip, using Apple’s MLX framework for on-device machine learning.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What Is Muse Spark Open, Exactly

Muse Spark Open is not the same product as Meta’s consumer-facing Muse app. That distinction matters because several outlets had to issue corrections after the initial wave of headlines implied the Muse chat assistant itself now runs on a Mac. Macworld pushed back directly on that framing, noting that the Muse app and its everyday models still run on Meta’s own servers in the cloud, while the model that actually runs locally is the separate Muse Spark Open release, as Macworld explained in its own breakdown of the confusion.

Meta released the weights for Muse Spark Open on October 15, 2026, making them available through Hugging Face and several API platforms in addition to the Apple-optimized local path. The model sits in the same broad category as other recent agentic releases this site has tracked, including Liquid AI’s on-device agent model, though Muse Spark Open is dramatically larger in raw parameter count. Meta’s own description, carried in its research announcement, calls it “the first model in its class” that can run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework, a claim repeated nearly verbatim across the day’s coverage.

Why Apple and Meta Are Cooperating Now

Apple and Meta have rarely appeared on the same side of a product announcement. Apple’s App Tracking Transparency rules cut into Meta’s ad business for years, and the two companies have sparred publicly over data practices, app store fees, and competing visions for how AI assistants should work on a phone. The October 15 collaboration does not erase that history. It is a narrower, hardware-specific arrangement: Apple wants proof that Mac Studio desktops, not just cloud GPU racks, can handle frontier-scale AI workloads, and Meta wants its open-weight models running in as many places as possible, including machines it does not own or operate.

Meta chief executive Mark Zuckerberg announced the Mac compatibility directly, writing that “Muse Spark Open, weights releasing today, runs great on new Mac Studios thanks to work with @johnternus and Apple.” That post, which The Verge captured in its coverage of the exchange, drew a public reply from Apple’s John Ternus, who wrote that it was “amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk,” adding that “Mac Studio keeps pushing the boundaries of on-device AI.” The public back-and-forth between two executives who rarely interact so directly is itself part of the story: it signals that this is a deliberate joint marketing moment, not an unsanctioned developer workaround.

A separate Meta statement, cited by MacDailyNews in its writeup of the announcement, put it more plainly: the company said it “worked with Apple to bring leading agent capabilities to your desktop for the first time on the Apple M5 Ultra Mac Studios.” Neither company has described this as a formal, long-term partnership agreement. CNBC and other outlets have been careful to frame it as a collaboration around a specific model and a specific chip generation, not a broader alliance covering future products.

Inside the 816 Billion Parameter Architecture

The headline parameter count, 816 billion total with 42 billion active, places Muse Spark Open in a mixture-of-experts style design, where only a fraction of the full network activates for any given token. That architecture is precisely why local deployment is even plausible. A dense model with 816 billion active parameters would be far too slow and memory-hungry for a desktop setup. By activating roughly 5 percent of its total weights per inference step, Muse Spark Open can lean on Apple’s unified memory architecture instead of requiring a rack of data center GPUs.

The model’s 1 million token context window is large enough to hold entire codebases or lengthy multi-document research sets in a single session, which lines up with Meta’s stated target uses: coding, software engineering, long-horizon tasks, and agentic workflows that require the model to track state across many steps. Native multimodal support means the model can take in more than plain text, though Meta’s public materials have focused their examples on coding and reasoning rather than image or audio tasks.

Specification	Muse Spark Open detail
Total parameters	816 billion
Active parameters	42 billion
Context window	1 million tokens
Modality	Natively multimodal
Release date	October 15, 2026
Minimum local hardware	Two Apple M5 Ultra Mac Studio units
Unified memory per unit	Up to 512 GB
Combined memory, two units	Up to 1 TB
Local framework	Apple MLX
Primary stated uses	Coding, agentic workflows, multi-step reasoning
Other distribution channels	Hugging Face and multiple API platforms

The Hardware Reality: Two Mac Studios and a Five-Figure Bill

Running Muse Spark Open at home is not a casual proposition. Meta’s own guidance states that the model can run on a desktop setup using two Apple M5 Ultra Mac Studios with the latest MLX framework, and each of those machines can supply up to 512 GB of unified memory, for a combined 1 terabyte across the pair. CNBC reported that a Mac Studio configured with the required M5 Ultra chip and 256 GB of memory costs at least $10,799, and also noted that customers can pair anywhere from two to four Mac Studio units depending on the workload.

Multiply that reported per-unit price by two and the minimum hardware outlay for Meta’s own documented configuration lands at roughly $21,598. Neither Apple nor Meta has published an official bundled price for the full two-machine setup, and the $10,799 figure CNBC cited applies to a specific 256 GB configuration rather than the 512 GB-per-unit tier Meta references for the full 1 TB combined memory pool. In practice, a buyer chasing Meta’s documented maximum-memory configuration should expect to pay more than $21,598 once they step up from 256 GB to 512 GB units, though exact Apple pricing at that tier was not confirmed in the day’s reporting.

That price tag puts Muse Spark Open’s “local” deployment firmly in prosumer and small-studio territory rather than consumer reach. It is a meaningfully different audience than the one targeted by cheaper on-device hardware pushes elsewhere in the industry, including NVIDIA’s RTX Spark desktop AI systems, which start at a fraction of the cost of a dual Mac Studio rig.

How Apple’s MLX Framework Makes This Possible

MLX is Apple’s own machine learning framework, built specifically to take advantage of the unified memory architecture found in Apple Silicon. Unlike traditional GPU setups where a model’s weights have to be shuttled between separate CPU and GPU memory pools, Apple’s unified memory lets the M5 Ultra’s integrated GPU address the same memory space the CPU uses, up to 512 GB per Mac Studio unit. That architecture is precisely why two machines, rather than a rack of discrete GPUs, can hold an entire 816 billion parameter model’s weights in a workable form.

Meta’s engineering teams reportedly tuned Muse Spark Open specifically to take advantage of that memory layout, which is the technical core of the “worked with Apple” language in both companies’ public statements. This is also the detail that separates Muse Spark Open from a simple quantized or shrunk-down model port. Meta did not just compress an existing checkpoint to fit on consumer hardware; it engineered the mixture-of-experts routing and memory footprint around what Apple’s chips can actually address.

Historical Context: From App Store Fights to Joint Announcements

It is worth remembering how adversarial the Apple-Meta relationship has been. Apple’s 2021 App Tracking Transparency rollout forced Meta to disclose, in a 2022 earnings call, that the change would cost the company roughly $10 billion in lost ad revenue for that year alone. Meta’s leadership spent years afterward publicly criticizing Apple’s platform control, while Apple executives defended the privacy changes as non-negotiable. Even as recently as this year’s debates over AI assistant integration, the two companies have pursued largely separate paths: Apple built its own on-device AI stack around Apple Intelligence, while Meta pushed its Muse assistant and Llama-lineage open models through its own cloud infrastructure and partner platforms.

Seen against that backdrop, a joint announcement with executives publicly complimenting each other’s engineering work is a real shift in tone, even if it is scoped narrowly to one model and one hardware line. It echoes a broader pattern across the AI industry this year, where competitors have increasingly cooperated on technical compatibility even while competing hard on products. This site has covered a similar dynamic with SGLang’s day-zero inference support for the same Muse Spark Open model, where an independent serving framework moved quickly to support Meta’s release the moment it shipped.

What This Means for Developers and Enterprise Buyers

For developers, the practical pitch is straightforward: a team that already owns or can justify two Mac Studio units gets a frontier-scale coding and reasoning model that never has to leave their office network. That matters for organizations with strict data residency rules, regulated industries, or simply a preference for not sending proprietary source code to a third-party API every time an agent runs a task. Meta’s framing of Muse Spark Open around coding, software engineering, and long-horizon agentic work suggests the company is targeting exactly these buyers rather than general consumers.

At the same time, the $21,598-and-up hardware floor means most individual developers will still reach for a hosted API rather than buying dedicated Mac Studio hardware. That tension, frontier capability versus accessible pricing, is the same one playing out across the broader on-device AI market this year, from Apple’s own idle-Mac-GPU cloud experiments to cheaper dedicated AI desktop hardware from chipmakers chasing the same workloads at a lower price point.

Enterprise infrastructure teams evaluating Muse Spark Open will also need to weigh it against the open-source tooling ecosystem growing around similar releases. Nvidia has been pushing in a related direction with its own open-source AI tooling push in Europe, giving buyers more than one open-weight path to evaluate before committing hardware budget to any single vendor’s recommended configuration.

Competitive Comparison: Local Mac Deployment vs Cloud-Hosted Frontier Models

Muse Spark Open’s local-on-Mac path is unusual because most frontier-scale models, including the current generations behind the leading hosted chat assistants, are designed to run exclusively on data center GPU clusters. Developers comparing everyday assistants, as covered in this site’s own look at how Claude, ChatGPT, and Gemini stack up for daily use, are almost always comparing cloud-hosted services with no local deployment option at all. Muse Spark Open sits in a different category: a model large enough to compete on raw capability claims, but distributed with weights anyone can download and, with enough Mac Studio hardware, run entirely offline.

The tradeoff is latency and throughput. Hosted frontier models benefit from data center-grade interconnects and purpose-built AI accelerators that outperform even a dual Mac Studio rig on raw tokens-per-second for heavy workloads. What Muse Spark Open offers instead is data control and a fixed hardware cost rather than a recurring API bill, a calculation that will appeal to some enterprise buyers and not others depending on usage volume.

Deployment path	Where it runs	Cost structure	Data control
Muse Spark Open (local)	Two or more Apple M5 Ultra Mac Studios	Upfront hardware, estimated $21,598 and up for two units	Fully on-premises
Muse Spark Open (hosted)	Hugging Face and third-party API platforms	Per-token API pricing	Third-party hosted
Hosted frontier chat assistants	Vendor data centers	Subscription or per-token API pricing	Third-party hosted
Smaller on-device agent models	Single consumer or prosumer device	One-time device cost, often under $3,000	Fully on-device

Market Impact for Apple and Meta

For Apple, the collaboration doubles as a validation story for the Mac Studio line and for MLX as a serious machine learning framework, not just a developer curiosity. Apple has spent two years positioning Apple Silicon’s unified memory as a selling point for AI workloads specifically, and having an outside company the size of Meta publicly engineer a flagship open-weight model around that architecture is a stronger endorsement than anything Apple’s own marketing could produce alone.

For Meta, the benefit is distribution and credibility in a market where Meta’s AI efforts are frequently compared unfavorably to closed, cloud-only competitors. By making Muse Spark Open runnable outside Meta’s own infrastructure, including on hardware Meta does not sell or control, the company reinforces its open-weight positioning at a moment when rivals are debating how open their own frontier models should be. Meta’s own characterization of the release as the most capable US open-weight model to date is a direct claim of leadership in that specific lane, separate from the broader frontier model race dominated by closed, proprietary systems.

Timeline: How the Announcement Unfolded

Date	Event
October 15, 2026	Meta publishes Muse Spark Open weights and research blog post
October 15, 2026	Mark Zuckerberg posts publicly about Mac Studio compatibility
October 15, 2026	Apple’s John Ternus replies publicly, praising Mac Studio performance
October 15, 2026	CNBC reports on the partnership and Mac Studio pricing
October 15 to 16, 2026	The Verge, MacRumors, TechCrunch, and MacDailyNews publish follow-up coverage
October 16, 2026	Macworld clarifies that the consumer Muse app itself still runs on Meta’s servers

What Industry Voices Are Saying

Beyond the two companies’ own statements, the reaction so far has centered on the unusual optics of Apple and Meta publicly celebrating each other’s engineering. Zuckerberg’s post crediting Apple’s John Ternus by name, and Ternus’s direct reply praising Mac Studio’s performance, read less like a routine press release and more like a coordinated moment both companies wanted documented in public. A separate internal Meta statement framed the work as bringing “leading agent capabilities to your desktop for the first time on the Apple M5 Ultra Mac Studios,” a line that frames the achievement around desktop-class agentic AI rather than raw benchmark scores.

That framing matters because it sets expectations for how both companies will market follow-up hardware and models. If Apple and Meta continue this pattern with future Mac Studio generations or future Muse releases, it would mark a genuine shift from one-off cooperation to a recurring joint go-to-market motion, something neither company has signaled with any other competitor in the current AI cycle.

Predictions: Where This Goes From Here

• Expect Apple to lean harder on MLX and Mac Studio AI benchmarks in its next hardware cycle marketing, using Muse Spark Open as a reference workload.
• Expect Meta to pursue similar local-deployment partnerships with other hardware makers to avoid looking dependent on a single vendor’s chips.
• Expect continued public confusion between the consumer Muse app and Muse Spark Open, requiring more clarifying coverage like Macworld’s.
• Expect enterprise buyers with strict data residency needs to treat the dual Mac Studio configuration as a serious, if expensive, on-premises option for coding and agentic workloads.
• Expect competitors building cheaper on-device AI hardware to use the roughly $21,598-plus Mac Studio price floor as a direct selling point against Apple’s approach.

Frequently Asked Questions

Does the Muse app now run locally on my Mac?
No. The consumer Muse app and its everyday models continue to run on Meta’s servers. Only the separate Muse Spark Open model, a developer-facing open-weight release, can run locally on qualifying Mac Studio hardware.

How much hardware do I need to run Muse Spark Open locally?
Meta states the model can run on a desktop setup using two Apple M5 Ultra Mac Studio units with the latest MLX framework. Other reporting mentions configurations of two to four Mac Studios depending on the workload.

How much does that hardware cost?
CNBC reported that a Mac Studio with the required M5 Ultra chip and 256 GB of memory costs at least $10,799. Two units at that price point would run at least $21,598, though neither company has confirmed an official bundled price for the full two-machine setup Meta describes with 512 GB per unit.

How large is Muse Spark Open compared to other open-weight models?
Muse Spark Open has 816 billion total parameters with 42 billion active parameters per inference step, a 1 million token context window, and native multimodal support, according to Meta’s research announcement.

Is this a formal partnership between Apple and Meta?
Neither company has described it that way. Reporting characterizes it as a collaboration focused on optimizing one model for one hardware line using Apple’s MLX framework, not a broader or long-term alliance.

What is Muse Spark Open designed for?
Meta positions it for coding, software engineering, long-horizon multi-step tasks, and other agentic workflows rather than general consumer chat use.

Can I download Muse Spark Open without buying Mac Studio hardware?
Yes. Meta released the model’s weights on Hugging Face and made it available through several API platforms, independent of the Apple-optimized local deployment path.

Why did Apple and Meta work together on this instead of competing?
Apple gains a flagship proof point that Mac Studio and MLX can handle frontier-scale AI workloads, while Meta gains wider distribution for its open-weight model strategy on hardware it does not manufacture or control.

Related Coverage

• Meta Stock Jumps 4% on Muse Wonder Frontier Bet [2026]
• Meta Prices Muse Code at $200 to Undercut AI Rivals [2026]
• Anthropic's 2nd Exec Warns Open AI Is Riskier [2026]
• Apple's Siri AI Takes On Alexa+ in $299 HomeView [2026]
• OpenAI's 372 Math Problems Spark 'Mathpocalypse' [2026]