Meta released the model's weights for Muse Spark Open on Thursday — the most capable open-weight AI model developed in the United States — and simultaneously announced that two Apple Mac Studio desktops, linked over Thunderbolt 5, can run it entirely on-premises without a cloud connection. The Apple-Meta collaboration confirmed by both companies is the first public demonstration of Thunderbolt 5 Remote Direct Memory Access (RDMA) serving as a frontier-AI interconnect — a prosumer-grade alternative to the data-center networking that has, until now, been the only viable path to running a model at this scale.

Mark Zuckerberg announced the collaboration on X alongside a photograph of Muse Spark Open running on a Mac, and Apple CEO John Ternus responded publicly on X: "Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI."

Muse Spark Open: 816 Billion Parameters, One Million Tokens, Entirely On-Premises

Muse Spark Open carries 816 billion total parameters organized in a sparse mixture-of-experts (MoE) Transformer architecture, of which roughly 42 billion activate at inference time per token. That split is architecturally significant: the model's compute requirement at inference scales with the active parameter count (roughly equivalent to running a 42B dense model), while its memory requirement scales with the total parameter count — all expert sub-networks must remain loaded in accelerator memory simultaneously, because the routing decision for any given token can invoke any expert. This is why reaching 816B at the desktop level is a memory problem, not a compute problem, and why it took a cluster approach to solve it.

The model's one-million-token context window — approximately 750,000 words — is large enough to hold entire codebases, hour-long transcripts, or multi-document research sessions in a single inference call. It accepts text, images, video, and audio as inputs and produces text output, and Meta has specifically optimized it for coding and long-horizon agentic tasks: the official blog cites completing DaVinci Resolve video-editing workflows end to end and shipping real iOS changes to the open-source Ice Cubes Mastodon client as real-world validation tasks.

Meta is releasing Muse Spark Open under a Muse Model License that permits limited commercial use. Weights are available today on Hugging Face at the meta-models Muse Spark Open collection, as well as through Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

Thunderbolt 5, RDMA, and the Terabyte Memory Pool

Running a model of this scale outside a data center was, until recently, essentially impossible. The memory footprint alone — holding all of the model's expert sub-networks at inference weight precision — requires more RAM than any single commercially available machine carries. Apple's answer is to make two machines appear, to the application, as one.

The validated baseline configuration pairs two M5 Ultra Mac Studios, each equipped with 512 GB of unified memory, for a combined pool of one terabyte. A four-machine configuration using 256 GB units each achieves the same total capacity at the cost of additional inter-node latency. The M5 Ultra — a quad-die fusion of four M5 chips presenting as a single logical processor — delivers 1.2 TB/s memory bandwidth per machine — which means each node can feed its neural engines the model's weight data at rates that keep inference interactive rather than agonizingly slow.

The critical enabling technology is Thunderbolt 5 combined with RDMA. Apple first shipped RDMA over Thunderbolt 5 in macOS Tahoe 26.2, released in December 2025, and formalized the hardware capability with the M5 Ultra Mac Studio's August 2026 announcement. Thunderbolt 5 delivers interconnect bandwidth of up to 120 Gb/s.

RDMA allows one Mac Studio's neural-network accelerators to read and write directly into the peer system's unified memory — bypassing the CPU and OS kernel on both ends — at sub-millisecond latency. The result is that the two machines' separate 512 GB pools merge into a single 1 TB address space from the perspective of the ML framework running on top.

Apple demonstrated the principle at WWDC 2026, showing a four-node M3 Ultra cluster generating tokens for a 27-billion-parameter model at nearly three times the rate of a single machine. Apple's August 2026 Mac Studio announcement confirmed that a four-system cluster delivers up to 3x faster AI inference than a single system. Muse Spark Open's launch is the first production application of the same architecture at frontier scale.

Apple's MLX: How Distributed Inference Reaches Your Desk

The software layer that makes this possible is Apple's MLX framework — an open-source array library built by Apple ML Research specifically for the unified memory architecture of Apple Silicon. MLX has supported distributed pipeline parallelism across a Thunderbolt-linked cluster since mid-2026, and for Muse Spark Open's launch, Meta tuned the model weights specifically for MLX execution and Apple published a dedicated adapter.

The adapter handles layer distribution automatically. In a two-machine cluster, Muse Spark Open's transformer layers are split across the two nodes; the RDMA interconnect carries the activation tensors between layers from the first machine to the second, without involving either machine's CPU in the transfer. From the developer's perspective, launching distributed inference requires a single mlx.launch command pointing at a hostfile listing the two machines' addresses — no MPI configuration, no driver management, no manual parallelism code. The MLX adapter for Mac Studio is available now.

The low-level infrastructure making that single command possible is JACCL — an open-source collective communications library built by Apple, named after Jack Beasley, who led the Thunderbolt 5 RDMA development effort, and Angelos Katharopoulos from the MLX team. Introduced at WWDC 2026, JACCL's name is a deliberate nod to NVIDIA's NCCL (NVIDIA Collective Communications Library). JACCL handles the AllReduce, AllGather, and related collective communications primitives that distributed inference requires, and MLX calls it automatically without requiring developer involvement.

Meta also contributed GPU kernel optimizations for the M5 Ultra's Apple GPU architecture, with the official launch blog reporting performance gains of 1.03 to 3.95 times across key compute kernels.

An Unlikely Alliance: How Apple and Meta Built This Together

The collaboration carries weight beyond its technical details. Apple and Meta spent years as public adversaries over privacy. Meta CEO Mark Zuckerberg spent much of 2021 and 2022 openly criticizing Apple's App Tracking Transparency (ATT) framework, which curtailed Meta's advertising data pipeline on iOS and cost the company an estimated billions in revenue. That history makes Thursday's joint announcement — and Zuckerberg's personal celebration of it on X — a notable détente in Silicon Valley's most visible technology rivalry.

The person most credited with bridging the two companies is Daniel Gross, now a member of Meta's Superintelligence Labs unit. Before joining Meta in July 2025, Gross co-founded Safe Superintelligence with Ilya Sutskever — a company whose founding premise was that safety-first research should precede commercial AI deployment.

Before that, Gross served as Apple's Director of Machine Learning for several years, where he contributed to Siri and Apple's machine learning infrastructure during the period that would eventually produce Apple Silicon and the unified memory architecture now at the center of this deployment. Alexandr Wang, Meta's Chief AI Officer, confirmed publicly that the two companies had been working together for months before Thursday's release, and credited Gross with playing "a really big role in kind of getting the two companies to work nicely together."

Wang has been among the most vocal advocates for the position that open-weight model development is essential to US competitiveness in AI — a position he argued before the US Senate and that underpins Meta's decision to release Muse Spark Open's weights publicly rather than keep them proprietary. That policy stance, combined with Apple's hardware and software contributions, makes this collaboration as much an alignment of strategic interests as a technical one.

Meta's internal commitment to the platform reinforced that signal: the company had already deployed hundreds of Mac machines internally for its Superintelligence team before the public launch, providing both a testing ground and a reputational stake in making the hardware work.

What It Costs, and What It Enables

The engineering achievement carries a substantial price tag. A Mac Studio with the M5 Ultra and 256 GB of unified memory currently starts at $10,799. The 512 GB configuration required to run Muse Spark Open across two machines in the validated baseline is expected to ship later this month and will carry a higher premium; that pricing has not yet been announced as of Thursday. A minimally viable two-machine cluster, using two 512 GB units, will cost well in excess of $20,000 before displays or storage.

For organizations or developers for whom that price point makes sense, the calculus combines two distinct advantages. Cloud inference on a frontier-class model at scale carries per-token costs that compound quickly at high usage, and it routes every prompt, file, and piece of context through a third-party provider's infrastructure. Local inference eliminates both the recurring cost and the data-routing dependency. For law firms, healthcare systems, defense contractors, financial institutions, and any organization handling sensitive or regulated data, the second advantage may outweigh the first.

The competitive implications extend beyond this single deployment. For Nvidia, whose NVLink and InfiniBand interconnects define the high-performance end of AI-specific clustering, Thursday's announcement is a proof of concept that Thunderbolt 5 RDMA can serve as a prosumer-grade alternative — not at data-center scale, but competently at the frontier desktop tier.

Community benchmarks from earlier in 2026 showed four-node Mac Studio clusters running the Kimi K2 Thinking model (a one-trillion-parameter, 32-billion-active MoE model) at approximately 25 tokens per second, with total cluster power draw measured in community tests at approximately 450 to 600 watts. Those results suggest that two to four Mac Studios represent a materially different point on the cost-per-capability-per-watt curve than a comparably capable GPU cluster.

What Apple and Meta have built is also reusable. The Thunderbolt 5 RDMA + MLX pipeline parallelism stack is not Muse Spark Open-specific — it is a general-purpose prosumer distributed inference architecture. Any open-weight frontier model whose total weight set fits a Thunderbolt-linked Mac Studio cluster's combined unified memory can, in principle, run on this same infrastructure. The barrier to desktop-scale frontier inference has not been lowered for one model. It has been lowered for the category.

Availability

Muse Spark Open is available today. Model weights can be downloaded from Hugging Face at the meta-models Muse Spark Open collection, and are also available through Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. The MLX adapter for Mac Studio clusters is also available now.

Developers who do not yet have the required hardware can evaluate the model through hosted API providers: Together AI, Fireworks AI, Baseten, Modal, Groq, Nebius, and OpenRouter all offer Muse Spark Open inference as of Thursday. The Muse Model License permits limited commercial use; organizations with high-volume commercial requirements should review the license terms before deployment.

Frequently Asked Questions

What is Muse Spark Open, and how is it different from other large AI models?

Muse Spark Open is Meta's largest publicly released open-weight model — 816 billion total parameters organized in a mixture-of-experts architecture, with 42 billion parameters active per token at inference. It accepts text, images, video, and audio as inputs and is optimized for coding and long-horizon agentic tasks with a one-million-token context window.

What sets it apart is the combination of its scale (the largest open-weight model developed in the United States), its access model (weights freely downloadable), and its validated path to on-premises desktop inference via Apple's Mac Studio cluster — something no previous frontier-scale model has achieved outside a data center. Full details are in Meta's official Muse Spark Open blog.

Why does running a model this large require two computers instead of one?

The mixture-of-experts architecture forces all of the model's expert sub-networks — spanning the full 816-billion-parameter weight space — to remain simultaneously loaded in accelerator memory, because any given token may route to any expert. At roughly two bytes per parameter (bfloat16 precision), the full weight set requires approximately 1.5 to 1.6 terabytes of memory in theory; in practice, Apple's MLX implementation with quantization fits within one terabyte. No single commercially available machine carries that much RAM. Two M5 Ultra Mac Studios, each with 512 GB of unified memory, provide the combined capacity — and Thunderbolt 5 RDMA allows both machines' memory pools to function as a single address space from the application's perspective.

Does this mean any future large open-weight model will run on a Mac Studio cluster?

The Thunderbolt 5 RDMA and Apple MLX pipeline parallelism stack that powers Muse Spark Open's deployment is not specific to this model — it is a general infrastructure layer. Any open-weight model whose full weight set fits within the combined unified memory of a Thunderbolt-linked cluster could, in principle, run on the same architecture.

Apple and Meta have effectively established a reusable prosumer distributed inference platform. Whether a given future model fits depends on its total parameter count, the precision its weights are stored at, and how many Mac Studios are in the cluster; the current validated maximum for a two-machine setup is one terabyte. The MLX adapter documentation covers how to configure a cluster for other models.

Should I buy the hardware now, or wait for the 512 GB Mac Studio to ship?

The 512 GB M5 Ultra Mac Studio configuration — required for the two-machine baseline — has not yet shipped as of Thursday's announcement; Apple says it will ship in late October at a price it has not yet disclosed. The 256 GB configuration, which starts at $10,799, is available now; a four-machine cluster of 256 GB units achieves the same 1 TB total capacity as the two-machine 512 GB setup, at the cost of additional network hops. Developers who want to evaluate the model before committing to hardware can use any of the seven API providers offering hosted inference today.