Apple-Meta Muse Spark Open: Local AI and the Chip Supply Chain

1. What happened

On October 15, 2026, Meta released an open-weight large language model called Muse Spark Open and, in the same announcement, disclosed a technical collaboration with Apple. The partnership is unusual because the two companies have spent most of the past decade as adversaries rather than collaborators.

According to the announcement, Muse Spark Open can run locally on two Mac Studio systems equipped with Apple's M5 Ultra chip and 512 GB of unified memory each. Working together, the two machines behave like a single system with roughly one terabyte of unified memory. The inference stack relies on MLX, Apple's machine-learning framework for Apple silicon.

Meta chief executive Mark Zuckerberg publicly thanked Apple chief executive Terrence and the Apple team in a social-media post. Terrence replied that running a model of Muse Spark Open's scale and specification so efficiently on the desktop was remarkable. The exchange was brief, but it carried symbolic weight: the two firms were publicly acknowledging a working relationship.

The model itself is substantial. Muse Spark Open has 816 billion total parameters, uses a mixture-of-experts design, and activates 42 billion parameters per token. It supports a context window of up to one million tokens and is positioned for coding, agentic tasks, multi-step reasoning, and software engineering workloads. Meta said the weights are available on Amazon, Google, and Microsoft cloud platforms, and that major inference frameworks support the model.

The cooperation did not emerge from nowhere. Daniel Gross, who founded the AI startup Cue, sold it to Apple in 2013 and later led parts of Apple's artificial-intelligence organization before leaving for Meta, was described by Meta AI head Alexandr Wang as a factor in bringing the two sides together. Wang said most of the historical hostility between the companies predates his arrival. Online observers joked that the parents had reconciled.

The timing matters. Meta released a stronger closed model, Muse Wonder, earlier in the same week. By open-sourcing the lower tier and keeping the frontier model proprietary, Meta is drawing a deliberate line between ecosystem expansion and competitive advantage.

2. Why it matters

The story is easy to read as a celebrity-tech drama: two proud companies, years of bad blood, a surprise handshake. That reading is not wrong, but it misses the more consequential signal for anyone who buys, sells, or distributes electronic components.

Local inference at the scale described here changes the demand profile for hardware. Cloud inference concentrates demand in data-center accelerators, high-bandwidth memory, and the power and cooling infrastructure that surrounds them. Local inference distributes demand differently: toward high-memory unified architectures, compact packaging, and the power-management and connectivity components that sit around a desktop-class system.

Apple's unified memory architecture is the technical hinge. In a conventional discrete-GPU system, model weights must move across a PCIe bus between system memory and graphics memory, and memory capacity is bounded by what fits on the accelerator. In a unified architecture, the CPU, GPU, and neural engines share one pool of memory. That pool can be large, and the model does not need to be sharded across separate memory spaces in the same way. Two machines linked together extend that pool further.

This is why the partnership is more than a marketing moment for Apple. It is a demonstration that a very large model can leave the cloud and run on hardware a developer or a small team can own. That demonstration has direct implications for the memory market, for advanced packaging, and for the component ecosystem around high-performance desktops and workstations.

There is also a strategic dimension for Meta. Open-weight models let Meta push its architecture into developer workflows without operating every inference endpoint itself. If Muse Spark Open runs on a Mac Studio in a studio apartment, Meta has extended its reach without adding a data-center bill. The trade-off is that Meta surrenders some control over how the model is used and monetized.

For Apple, the partnership reinforces a narrative the company has struggled to make convincing: that its silicon and its software stack are serious AI infrastructure, not just consumer conveniences. The company has faced persistent criticism that its assistant and on-device intelligence lag the leading labs. A credible local-inference demonstration is one way to answer that criticism without depending entirely on a third-party cloud model.

3. The rivalry in context

To understand why the handshake is notable, it helps to recall how strained the relationship had become. Apple and Meta are two of the largest and most valuable technology companies in the world, and their business models have pulled them in opposite directions on privacy and data.

Apple has positioned itself as a privacy-first hardware company, selling devices and services while arguing that user data should stay on the device. Meta has built an advertising business on detailed user profiles and cross-site tracking. Those models are structurally in tension, and the tension has repeatedly turned into public conflict.

The 2018 Cambridge Analytica scandal brought the differences into the open. Apple's chief executive Tim Cook criticized the handling of user data, and Meta's Mark Zuckerberg pushed back, arguing that a hardware company should not lecture an advertising company. The exchange hardened the public perception of the two firms as adversaries.

The 2021 App Tracking Transparency change turned the philosophical disagreement into a financial one. Apple required apps to ask users for permission before tracking them across other apps and websites. Many users declined. Meta estimated that the change would reduce its 2022 revenue by roughly $10 billion, and the company spent months arguing that the policy harmed small businesses that relied on targeted advertising. The relationship did not recover in the years that followed.

Against that backdrop, a technical collaboration is a genuine thaw. It does not erase the strategic conflict, but it shows that the two companies can cooperate when the technical and business incentives align. The involvement of Daniel Gross, who moved from Apple to Meta, gives the cooperation a human thread: a relationship that had been defined by public argument was partly rebuilt by people who had worked on both sides.

The broader industry context also matters. The AI race has forced every large technology company to reconsider old rivalries. Apple needs capable models; Meta needs distribution for its open-weight strategy. The incentives that once kept them apart have, at least on this narrow point, been outweighed by the incentives to work together.

It is worth being precise about what the thaw does and does not mean. The two companies remain competitors in devices, services, and advertising. Apple still sells the hardware on which Meta's apps run, and Meta still depends on platform rules that Apple sets. The cooperation is narrow: a technical demonstration on Apple silicon. It does not resolve the privacy conflict, and it does not mean the two companies will collaborate broadly. Read it as a pragmatic alignment on one problem, not a strategic alliance.

That narrowness is itself instructive. Large technology companies rarely reverse course on strategy because of a single event. They make tactical accommodations when the cost of not cooperating exceeds the cost of cooperating. For Meta, the cost of not having a credible local-inference story was rising; for Apple, the cost of not having a serious AI demonstration was rising. The collaboration is the point at which those costs crossed. It is a signal about incentives, not about affection.

4. The technical background

4.1 Apple silicon and unified memory

Apple's M-series chips place the CPU, GPU, neural engine, and memory controller on a single package with a shared memory pool. The M5 Ultra, the configuration described in the announcement, extends this design to a very large memory footprint. The practical consequence for language models is that weights, activations, and the key-value cache can reside in one address space, which simplifies inference and reduces the overhead of moving tensors between separate memory domains.

Memory bandwidth is the other half of the story. A large model does not only need to fit in memory; it needs to be read quickly enough to keep the compute units fed. Unified memory removes the copy overhead, but the underlying bandwidth still limits how fast tokens can be generated. This is why local inference performance is often described in terms of tokens per second rather than raw compute, and why memory bandwidth is a more useful specification than a headline clock speed.

Unified memory is not a cure-all. Bandwidth still matters, and the memory must physically exist. A 512 GB configuration is expensive because high-capacity memory packages are expensive and because the packaging that connects them to the compute die is itself a constrained resource. This is where the chip supply chain becomes relevant: the ability to build a desktop system with half a terabyte of unified memory depends on advanced packaging capacity and on the supply of high-density memory components.

The two-machine configuration is also technically interesting. Linking two systems to behave like one with roughly a terabyte of memory requires a fast interconnect and software that can coordinate inference across the boundary. The announcement frames this as a way to run models that would otherwise not fit in a single machine. In practice, cross-machine inference introduces latency and synchronization costs, so it is likely to be used for batch or offline workloads more than for latency-sensitive interactive use.

It is worth contrasting this with the data-center approach. A cloud inference cluster distributes a model across many accelerators connected by high-speed fabrics, with memory pooled and shared. The Apple-Meta demonstration is a scaled-down version of the same idea, using two machines instead of a rack. The economics and the failure modes differ, but the underlying problem, keeping a large model resident in memory, is the same.

4.2 MLX and the software stack

MLX is Apple's array and machine-learning framework, designed to take advantage of unified memory on Apple silicon. It allows arrays to be shared between CPU and GPU without explicit copies, which is exactly the property that makes large-model local inference feasible on a Mac. The choice of MLX in the announcement is a signal that the inference path is optimized for Apple hardware rather than ported from a data-center stack.

The software stack matters as much as the silicon. A model with hundreds of billions of parameters requires efficient quantization, memory management, and kernel implementations to run at acceptable speed on a desktop. The fact that Meta and Apple are presenting a working configuration suggests that the software maturity has reached a point where the demonstration is credible, not merely theoretical.

4.3 Mixture of experts and active parameters

Muse Spark Open uses a mixture-of-experts architecture. In a dense model, every parameter is used for every token. In a mixture-of-experts model, the network contains many specialized sub-networks, and a routing mechanism selects a subset for each token. The result is that the total parameter count can be very large while the number of parameters actually used per token, the active parameter count, stays smaller.

The announced figures illustrate the point: 816 billion total parameters, 42 billion active. The total determines how much memory is needed to hold the model; the active count determines, roughly, how much compute is used per token. This is why a model of this size can run on a desktop at all. It is also why memory capacity, rather than raw compute, becomes the binding constraint for local deployment.

The one-million-token context window adds another dimension. A long context increases the size of the key-value cache, which competes with model weights for memory. On a unified-memory system, that competition is visible and direct. The ability to run long-context workloads locally therefore depends on having memory headroom beyond the weights themselves.

4.4 Local inference versus cloud inference

Cloud inference offers elasticity, centralized management, and access to the largest models. Local inference offers privacy, predictability, and independence from network conditions. The Apple-Meta demonstration sits at the intersection: a model large enough to be useful, running on hardware that a developer can own, with data that does not leave the machine.

The privacy argument is not merely rhetorical. For organizations that handle sensitive code, regulated data, or proprietary designs, the ability to run a capable model locally removes a class of compliance questions. For individual developers, it removes a recurring cost and a dependency on external availability. These are the reasons local inference has persisted as a niche even as cloud inference dominated headlines.

4.5 The economics of local inference

The economics are less obvious than the technical story. On a per-token basis, cloud inference is usually cheaper because providers amortize hardware across many customers and buy at scale. Local inference shifts the cost model: the buyer pays upfront for hardware that sits idle much of the time. The break-even point depends on utilization, on the value of privacy, and on whether the buyer would otherwise pay for cloud tokens at retail rates.

For a developer running a capable model intermittently, the math can favor ownership. For an organization running continuous inference at high volume, cloud may remain cheaper unless privacy or data residency makes local deployment mandatory. The Apple-Meta demonstration does not settle this calculation. It expands the set of workloads where local deployment is technically possible, and leaves the economics to the buyer.

There is also a resilience argument that does not appear in spreadsheets. A local system does not depend on a provider's availability, on network conditions, or on a pricing change. For workloads where interruption is costly, that resilience has value that is hard to quantify but real. The same argument applies to supply chains: a component buyer who depends on a single source understands the value of an alternative, even if the alternative costs more per unit.

The cost curve also matters over time. Hardware that is expensive today tends to become cheaper as memory density improves and packaging capacity expands. If the local-inference category grows, the cost of the demonstrated configuration should fall, which would widen the set of buyers for whom ownership makes sense. The trajectory, not the current price, is the more interesting signal.

5. The AI hardware supply chain

The announcement is a useful lens on how AI demand is reshaping the component supply chain. Three layers matter: the logic die, the memory, and the packaging that connects them.

5.1 The logic die

Apple designs its own silicon and relies on foundry partners for manufacturing. The M5 Ultra is an advanced-node product, which means its supply is tied to the leading-edge capacity of the foundry that produces it. Every large customer that commits to advanced-node capacity competes for the same finite manufacturing slots. When a company like Apple signals a new AI-heavy workload for its silicon, it is also signaling demand for the advanced-node capacity behind it.

5.2 The memory

High-capacity unified memory depends on high-density DRAM and on the ability to package it close to the compute. The memory market has been volatile for years, and AI demand has added a new source of pressure. The Apple-Meta demonstration, if it becomes a category, would add another demand vector for high-capacity memory components, distinct from the data-center HBM demand that has dominated recent headlines.

5.3 The packaging

Advanced packaging is the quiet bottleneck. Connecting a large compute die to a large memory pool requires sophisticated packaging technologies, and packaging capacity has been tight across the industry. The ability to build a desktop with half a terabyte of unified memory is, in part, a packaging question. Buyers who track only logic die supply and ignore packaging capacity miss half the constraint.

5.4 The memory market context

The memory market has its own dynamics. High-bandwidth memory for data centers has absorbed enormous capacity in recent years, and the suppliers that make it have been allocating capital toward that demand. A new class of demand from high-memory desktop systems would compete for some of the same underlying capacity, even if the specific products differ. For buyers, the practical implication is that memory availability is not a fixed backdrop; it is a variable that moves with every new demand vector.

The volatility cuts both ways. Periods of oversupply can make memory cheap and abundant; periods of tight supply can push lead times out and prices up. A procurement strategy that assumes one regime persists indefinitely is fragile. The Apple-Meta demonstration, if it becomes a category, would add another source of demand to an already cyclical market, which argues for keeping visibility into memory supply rather than treating it as a commodity afterthought.

There is also a specification dimension. High-capacity unified memory is not the same product as data-center HBM, and the two are not freely substitutable. A supplier tooled for one may not be tooled for the other. The supply-chain implication is that capacity in one segment does not automatically relieve pressure in another, which is why segment-level visibility matters for buyers who straddle both.

6. The numbers

The announcement contains a set of figures worth putting in one place. The table below summarizes the model and hardware parameters as stated, alongside context figures that help interpret them.

The relationship between total and active parameters is the single most important number pair in the announcement. It explains why a model that would ordinarily require a rack of accelerators can be presented as a desktop workload. It also explains why the memory bill, not the compute bill, dominates the local deployment cost. If the model needed 816 billion active parameters, the story would be very different.

The cost figures deserve scrutiny. A 256 GB Mac Studio above $10,000 is already a serious purchase for an individual. The 512 GB configuration, which the demonstration actually uses, is not broadly available, which means the demonstrated configuration is closer to a reference design than to an off-the-shelf product. Buyers should treat the announcement as a capability demonstration and watch for commercial availability before planning procurement around it.

7. Who is affected

7.1 Apple

Apple gains a credible AI-infrastructure story for its high-end Mac line. The company has been criticized for lagging in assistant quality and for relying on partners for frontier capabilities. A local-inference demonstration with a model of this scale supports the argument that Apple silicon is a genuine AI platform. The risk is that the demonstration remains a niche capability if the hardware is too expensive or too scarce to reach a broad developer audience.

7.2 Meta

Meta extends the reach of its open-weight strategy. By making a capable model runnable on developer-owned hardware, Meta lowers the barrier to experimentation and builds goodwill in the developer community. The decision to open the lower tier while keeping Muse Wonder closed is a calculated balance: open enough to seed an ecosystem, closed enough to preserve a frontier advantage.

7.3 Semiconductor and memory suppliers

The most direct supply-chain effect runs through memory and packaging. High-capacity unified memory configurations depend on advanced packaging and on the supply of high-density DRAM. If local inference at this scale becomes a durable category, demand for these inputs rises, and the constraint shifts from compute dies to memory capacity and packaging capacity. Suppliers of power-management, clocking, and interface components around high-performance desktops also see indirect demand.

7.4 Cloud providers

Amazon, Google, and Microsoft host the weights, which means the cloud providers still participate in distribution even as the model runs locally. This is a hedge for Meta and a reminder that local and cloud inference are complements as much as substitutes. The cloud providers retain the training, fine-tuning, and burst-inference workloads even if some inference moves to the edge.

7.5 Developers and enterprises

Developers gain an option that did not exist at this scale before: a capable model on hardware they control. Enterprises with data-sovereignty requirements gain a path to local deployment without building a data center. The cost and hardware availability remain the gating factors.

8. Apple's AI strategy and its partners

The Meta collaboration sits within a broader and somewhat unsettled Apple AI strategy. Apple has pursued a multi-partner approach rather than relying on a single model provider.

In 2024, Apple announced a partnership with OpenAI to bring ChatGPT into Siri. The integration was widely discussed but, according to OpenAI, underperformed relative to expectations. In July 2026, Apple sued OpenAI, alleging trade-secret misappropriation, which complicated the relationship. Earlier in 2026, Apple reached an agreement with Google to bring Gemini into Apple Intelligence. The result is a portfolio of partners rather than a single dependency.

The criticism has not gone away. Apple's new Siri, released in the prior month, was widely judged to lag the products of leading AI labs. The Meta collaboration gives Apple a different kind of answer: not a better cloud assistant, but a demonstration that its own hardware can run a serious model locally. Whether that answers the criticism depends on how broadly the capability reaches.

Meta's strategy is similarly layered. The company has invested heavily in open-weight models, releasing capable models for developers while keeping its strongest models closed. Muse Spark Open and Muse Wonder are the latest expression of that split. The open model seeds adoption; the closed model preserves an edge. The approach is not unique to Meta, but Meta has been among the most aggressive in pursuing it.

8.3 The open-weight strategy in context

Open-weight models occupy a specific niche in the AI landscape. They are not a charity; they are a distribution strategy. By releasing weights that anyone can run, a company can seed an ecosystem, attract developers, and shape how a model is used. The trade-off is that the company cannot fully control downstream use, and it risks eroding the value of its own closed models.

The layered approach attempts to manage that trade-off. Release a capable but not frontier model as open weights, and the ecosystem grows without giving away the leading edge. Keep the frontier model closed, and the company retains a commercial moat. The gap between the two tiers is the strategy. If the open model is too strong, the closed model loses value; if it is too weak, developers ignore it. Muse Spark Open sits on the capable side of that line.

For the hardware ecosystem, the open-weight strategy is a demand multiplier. Every open-weight model that can run on developer-owned hardware is a reason to buy that hardware. The more capable the open model, the stronger the pull. This is why the combination of an open-weight model and a credible local-inference demonstration matters more than either alone.

The strategy also has a competitive dimension. When one company releases capable open weights, others face pressure to respond, either by releasing their own open models or by differentiating on proprietary capabilities. The result is a dynamic equilibrium in which the frontier stays closed and the capable tier becomes increasingly open. Buyers of AI hardware benefit from this dynamic because it widens the set of models that can run on hardware they own.

9. Scenarios

Three scenarios frame the range of outcomes over the next twelve to twenty-four months.

9.1 Optimistic: local AI becomes a durable category

In the optimistic case, the demonstration is the start of a durable shift. High-memory desktop systems become a recognized AI workstation category, developers adopt local inference for privacy-sensitive and offline workloads, and demand for high-capacity memory and advanced packaging rises. Apple expands its high-memory configurations, and the component ecosystem around AI workstations grows. The trigger conditions are a commercial 512 GB configuration, a mature software stack, and visible enterprise adoption.

9.2 Neutral: a valuable niche

In the neutral case, local inference remains a valuable niche for privacy-sensitive teams, offline environments, and developers who prefer to own their tooling. Cloud inference continues to dominate volume. The announcement is remembered as an important proof of concept rather than the start of a mass market. The trigger conditions are steady but modest adoption and continued cost barriers.

9.3 Pessimistic: cost and complexity limit adoption

In the pessimistic case, the hardware cost, the scarcity of high-memory configurations, and the operational complexity of cross-machine inference keep adoption small. The demonstration remains a showcase. The trigger conditions are slow commercial availability, high memory prices, and competing cloud offerings that close the privacy gap.

9.4 How to read the signposts

The scenarios are not predictions; they are frames for watching the next twelve to twenty-four months. The most useful signposts are concrete. Watch whether a 512 GB configuration reaches broad commercial availability. Watch whether the software stack matures enough that cross-machine inference becomes routine rather than experimental. Watch whether enterprises with data-sovereignty requirements adopt local deployment at scale. Watch memory pricing and packaging lead times. Each of these is observable, and each shifts the probability among the scenarios.

The mistake to avoid is treating a single announcement as a trend. A demonstration is a data point; a trend is a pattern of adoption. The Apple-Meta collaboration is a strong data point because of the scale of the model and the prominence of the companies, but it is still one point. The follow-through, in hardware availability and in real workloads, is what will determine whether it becomes a trend.

10. What buyers should do

For procurement teams and component buyers, the practical takeaways are about watching the right indicators rather than reacting to headlines.

• Watch memory capacity and pricing. The binding constraint for local large-model inference is memory, not compute. Track high-density DRAM availability and packaging capacity as leading indicators.
• Distinguish demonstration from availability. The 512 GB configuration used in the demonstration is not broadly available. Plan around commercially shipping configurations, not reference designs.
• Map the surrounding component demand. AI workstations pull demand for power management, clocking, interface, and thermal components. Distributors that serve these categories should watch the category's growth.
• Keep dual-sourcing discipline. If local AI hardware becomes a procurement category, the same lead-time and allocation risks that affect data-center components will apply.
• Do not over-index on a single vendor. The local-inference story involves Apple silicon today, but the broader trend toward edge inference is not exclusive to one architecture.

The practical conclusion for buyers is to treat local AI hardware as an emerging category worth monitoring, not an established one worth committing to. The technical capability is real, the economics are unresolved, and the supply chain is the variable that will determine how quickly the category matures. Keeping visibility into memory and packaging, while maintaining sourcing discipline, is the position that holds up across all three scenarios.

11. Timeline

• 2017: The last widely cited public cooperation between Apple and Meta.
• 2018: The Cambridge Analytica scandal strains relations; Apple's Tim Cook and Meta's Mark Zuckerberg trade public criticism.
• 2021: Apple introduces App Tracking Transparency, reducing ad targeting; Meta estimates a roughly $10 billion revenue impact for 2022.
• 2024: Apple partners with OpenAI to bring ChatGPT to Siri.
• 2026, July: Apple sues OpenAI, alleging trade-secret misappropriation.
• 2026: Apple reaches an agreement with Google to bring Gemini into Apple Intelligence.
• 2026, October 15: Meta releases Muse Spark Open and announces the Apple cooperation.
• 2026, same week: Meta releases the stronger closed model Muse Wonder.

12. FAQ

What is Muse Spark Open? An open-weight large language model from Meta with 816 billion total parameters, 42 billion active parameters, and a one-million-token context window, positioned for coding and agentic workloads.

How does it run on a desktop? It uses MLX on two M5 Ultra Mac Studio systems with 512 GB of unified memory each, combining to roughly one terabyte of memory.

Why is the partnership significant? Apple and Meta have been rivals for years. A technical collaboration signals a thaw and gives Apple a credible local-AI demonstration and Meta a broader open-weight footprint.

Is the model as capable as Meta's best? No. Meta describes Muse Spark Open as weaker than its closed Muse Wonder model, reflecting a deliberate tiered strategy.

What does this mean for chip supply? It highlights memory capacity and advanced packaging as strategic constraints for local large-model inference, with indirect demand for power, clocking, and interface components.

Can I buy the demonstrated hardware today? The 256 GB configuration is available above $10,000; the 512 GB configuration is not yet broadly available.

Does this replace cloud inference? No. It complements it, serving privacy-sensitive and offline workloads while cloud inference continues to dominate volume.

Why does Meta open-source some models and not others? Meta balances ecosystem expansion against preserving a frontier advantage, opening capable-but-not-frontier models while keeping its strongest model closed.

What is mixture of experts? An architecture that routes each token to a subset of specialized sub-networks, so total parameters can be large while active compute stays smaller.

Why does unified memory matter for AI? It lets the CPU, GPU, and other processors share one memory pool, so large models can stay resident without constant copying between separate memory spaces.

What is a key-value cache and why does it matter? The KV cache stores intermediate attention tensors so the model does not recompute them for every token. It grows with context length, so long contexts consume memory that would otherwise be available for weights.

Will this make cloud inference obsolete? No. Local and cloud inference serve different needs. Local deployment favors privacy and offline use; cloud favors elasticity and access to the largest models. Most organizations will use both.

What should component buyers watch? High-density DRAM availability, advanced packaging capacity, and the lead times for the power, clocking, and interface components that surround AI workstations.

Is the hardware configuration standard? The 256 GB configuration is commercially available; the 512 GB configuration used in the demonstration is not yet broadly available, so treat it as a reference design.

13. Glossary

Unified memory: A memory architecture in which the CPU, GPU, and other processors share a single pool of memory rather than separate memory spaces.

Mixture of experts (MoE): A model architecture that routes each token to a subset of specialized sub-networks, keeping active compute below the total parameter count.

Active parameters: The number of parameters actually used per token in a mixture-of-experts model.

Context window: The maximum number of tokens a model can consider at once.

MLX: Apple's machine-learning framework optimized for unified memory on Apple silicon.

Open-weight model: A model whose trained weights are released for others to run and modify.

Inference: The process of running a trained model to produce outputs.

KV cache: Stored key and value tensors that accelerate attention over long contexts, consuming memory proportional to context length.

Advanced packaging: Technologies that connect multiple dies and memory stacks into a single package, a key constraint for high-performance AI hardware.

Tokens per second: A common measure of inference speed, describing how many output tokens a model generates each second. It depends on memory bandwidth as much as on raw compute.

Quantization: Reducing the numerical precision of model weights to lower memory use and speed up inference, at some cost to accuracy.

14. Sources

• Meta's announcement of Muse Spark Open and the Apple collaboration, October 15, 2026 (Meta Newsroom).
• Public statements by Meta chief executive Mark Zuckerberg and Apple chief executive Terrence regarding the collaboration, October 2026 (Apple Newsroom).
• Coverage of the model specifications and hardware configuration by industry press, October 2026 (Meta AI Blog).

For component sourcing and BOM pricing across AI-adjacent hardware, see our BOM quote service and our analysis of AI optical modules sold out through 2029. Related supply-chain context is available in our four signals reshaping the 2026 chip supply chain.