Meta Builds the Models. Apple Sells the Rig.

On October 15, Meta released the weights for Muse Spark Open — an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token and a one-million-token context window. It is, by Meta's account, the most capable American open-weight model to date, built for coding and agentic work. The same day, Apple's machine learning research team published a technical guide to distributed inference on Mac Studio, showing exactly how to run the model across two or four M5 Ultra machines using the MLX framework over Thunderbolt 5 RDMA.

📖 Read the full version with charts and embedded sources on ComputeLeap →

View original post on X →

Neither company used the word "partnership." They didn't need to. The top comment on the Hacker News thread — which hit #4 at 355 points and 188 comments — said what everyone was thinking: "So, Meta will spend millions/billions on building open-weight models and then people will send Apple $20,000 to run them. Seems like a good deal for Apple."

That commenter nailed the structural story. This isn't a product launch. It's the emergence of a distribution alliance — the kind that reshapes an industry without anyone signing a contract. Meta makes the models. Apple sells the silicon. The incentives are self-reinforcing, and the implications go far beyond a single model release.

The Model: What Muse Spark Open Actually Is

Muse Spark Open is a mixture-of-experts model: 816 billion total parameters, with 256 experts and 8 active per token, yielding 42 billion active parameters per forward pass. It ships with native multimodality (text, image, and video understanding), a one-million-token context window, and it was explicitly designed for agentic and coding workflows.

The practical specs that matter for inference:

• Memory footprint: roughly 493 GB in NVFP4 quantization (the format Apple benchmarks use), or over 1.6 TB in native BF16
• Inference engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth
• Weight distribution: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models
• Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI

ℹ️ Meta calls Muse Spark Open "the most capable US open-weight model to date." That qualifier — "US" — is doing real work. Chinese labs Moonshot (Kimi K3) and Zhipu (GLM 5.3) still lead on some benchmarks, particularly terminal coding, where K3 scores 85.0 vs. Muse Spark Open's 76.4 on Terminal-Bench 2.1.

The agentic demos are where Meta made its pitch concrete. Running entirely on-device through Muse Code (Meta's agent framework), the model shipped real iOS changes to Ice Cubes — an open-source Mastodon client — including bug fixes, new features, and full builds. It also drove DaVinci Resolve, reviewing footage frame by frame and editing autonomously. These are vendor demonstrations, not independently verified benchmarks, but they show what Meta thinks the model's sweet spot is: long-horizon, tool-using workflows that benefit from massive context and local execution.

The Hardware: Apple's MLX Distributed Inference Benchmarks

Apple's technical post is the more interesting half of this story. It's not marketing — it's a benchmark paper with real numbers, published by Apple's ML research team.

The headline configuration: two M5 Ultra Mac Studios, each with 512 GB of unified memory and 1.2 TB/s of memory bandwidth, connected over dual Thunderbolt 5 cables using RDMA (Remote Direct Memory Access). Together, they provide roughly 1 TB of unified memory and 2.4 TB/s of aggregate bandwidth.

Here's what Apple measured for decode speed (tokens/sec):

Model	2x M5 (Pipeline)	2x M5 (Tensor)	4x M5 (Tensor)
Muse Spark Open (816B, NVFP4)	15.7	25.0	36.7
Mistral Medium 3.5 (128B, FP8)	7.7	13.7	23.4
Qwen 3.8 (27B, BF16)	19.4	33.3	50.9

Apple frames 36.7 tokens per second as "well above average human reading speed." That's true — but it's also a rhetorical move. Cloud inference endpoints routinely deliver hundreds of tokens per second. The Mac Studio's advantage isn't speed. It's sovereignty: zero tokens leave your building.

⚠️ Benchmark caveat: The HN community flagged that Apple is "sandbagging Qwen pretty heavily by benchmarking Meta's model at 4-bit, against Qwen only at native BF16." The comparison flatters Meta's model — a 4-bit 816B MoE against a full-precision 27B dense model. Apple chose the configurations that make their partner's model look best relative to the competition.

The distributed inference protocol uses jaccl over RDMA over Thunderbolt 5, requiring macOS 26.2 or later. Apple supports both pipeline parallelism (splitting by depth) and tensor parallelism (splitting by width). For Muse Spark Open, tensor parallelism wins handily on decode, while pipeline parallelism has a modest edge on very long prefills.

Time to first token tells the latency story. At 8,192-token context, the four-machine setup hits 4.8 seconds for Muse Spark Open. At 65,536 tokens, that stretches to 40.6 seconds — workable for batch processing, painful for interactive coding.

Peak power draw is roughly 300W per M5 Ultra — Apple describes the entire four-machine setup as drawing "a fraction of the power used by a common toaster." That framing is more telling than the number: Apple is positioning the Mac Studio as office-grade infrastructure, not data-center hardware.

The Alliance Nobody Announced

Here's where the analysis diverges from the press coverage.

Every news outlet reported this as "Meta and Apple collaborate on AI." That framing misses the point. This isn't a collaboration — it's a convergent-interest distribution stack, and the comparison that matters isn't to any AI partnership. It's to Wintel.

In the 1990s, Intel made the chips and Microsoft made the OS. Neither company needed a formal alliance. Intel's chips ran Windows best. Windows ran best on Intel chips. Every sale of one product reinforced demand for the other. The flywheel was structural, not contractual.

The Meta-Apple dynamic has the same shape:

• Meta's incentive: Every Mac Studio sold is another node that runs Muse models locally. More local inference means less dependence on cloud providers who might favor their own models. Meta doesn't sell inference — it sells the ecosystem that makes its models the default choice.
• Apple's incentive: Every frontier open-weight model that gets bigger makes unified memory more valuable. NVIDIA's VRAM wall (32 GB on RTX 5090, 128 GB on DGX Spark) becomes Apple's moat. The bigger the models, the fewer hardware options that can run them without offloading — and Apple's is the only desktop-class option with 512 GB of unified memory.

Alexandr Wang, Meta AI's chief, told Axios: "We're excited for mommy and daddy to be playing nice again." The quip got laughs. But the subtext is real — Meta and Apple have spent years fighting over privacy, app store fees, and data access. AI inference is the first domain where their interests genuinely align.

View original post on X →

Hugging Face CEO Clement Delangue was characteristically direct. Within an hour of the announcement, he posted a one-liner: llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K (on two mac studios) — demonstrating that the open ecosystem had already packaged the model for the Mac Studio hardware. The distribution pipeline is immediate. No app store review. No cloud provider approval. Weights on Hugging Face, GGUF quantization, llama serve, and you're running a frontier model on your desk.

View original post on X →

What the Community Is Saying

The convergence on this story is striking. Our intel tracking caught it across four independent source categories on the same day — YouTube AI channels, YouTube tech channels, Hacker News, and Substack newsletters — all framing the same structural thesis before any outlet published a think piece.

View on Hacker News →

The HN thread (355 points, 188 comments) is where the practitioner signal lives. Beyond the top comment about Apple's deal, the second-most-upvoted comment provides the real benchmark context:

ℹ️ "FWIW, single M5 Ultra with 256GiB runs DeepSeek 4.1 Flash 0910 just fine (~3200 tok/s prefill, ~45 tok/s). M5 Max w/ 128GiB runs GLM 5.3 Flash also fine." — HN commenter

This is the data point Apple's marketing didn't highlight: you don't need two $10K+ machines for useful local inference. A single M5 Ultra with 256 GB already runs competitive open-weight models at respectable speeds. The Muse Spark Open configuration is the prestige play — the "we can run nearly a trillion parameters on your desk" headline. The practical play is the single machine running 70B-120B models at 45+ tok/s.

Nathan Lambert's Interconnects newsletter lands the week's analytical frame: "American open-weight models took a step change back toward the frontier." He explicitly pairs his open-model analysis with "the Apple MLX story on HN," confirming that the analyst community sees the same convergence: the model story and the hardware story are the same story.

Contrarian Corner: Is This Actually an Alliance?

⚠️ The bear case deserves a fair hearing. Newsquawk argued this is "a porting exercise, not a commercial deal" — engineers got an existing model running on existing silicon. Meta optimized for MLX the same way it optimized for vLLM, TensorRT-LLM, and llama.cpp. Apple benefits from any software that sells hardware. Where's the alliance?

The counter-arguments are real:

Cost: Engadget estimates a four-machine setup at roughly $38,000. Even two 512 GB machines will likely exceed $20,000. For that money, you could buy years of API access to hosted Muse Spark Open on Groq or Together AI — at hundreds of tokens per second instead of 37.

Speed: 36.7 tok/s on four Mac Studios is respectable for local inference but mediocre for production workloads. A hosted endpoint running the same model on NVIDIA H200s will generate tokens 5-10x faster. If you're building a product that serves users, the Mac Studio is the wrong tool.

Benchmarks: The HN community correctly flagged that Apple cherry-picked quantization levels. Comparing Muse Spark Open at NVFP4 against Qwen at native BF16 is like benchmarking a turbocharged V6 against a naturally aspirated four-cylinder and declaring victory. The comparison flatters Meta's model on Apple's hardware.

The real test: Does this configuration actually compete with cloud APIs for the developers who would consider a $20K hardware purchase? The answer depends on workload shape — which brings us to the math.

The Math: When Does Local Beat Cloud?

Let's be concrete. Two 512 GB M5 Ultra Mac Studios at an estimated $20,000 total (Apple hasn't announced pricing for the 512 GB configuration, but the 256 GB model already exceeds $10,000):

Factor	Mac Studio (2x M5 Ultra)	Cloud API (Together AI)
Upfront cost	~$20,000	$0
Monthly amortized (3 years)	~$555/mo	Variable
Decode speed	25 tok/s (tensor parallel)	200-500+ tok/s
Marginal cost per token	$0 (electricity ~$15/mo)	~$0.60/M input, $1.80/M output
Privacy	Full local — zero data leaves your network	Provider sees all tokens
Availability	Always on, no rate limits	Subject to provider uptime

Break-even analysis: At Together AI's Muse Spark Open pricing, you'd need to generate roughly 11 million output tokens per month to match the $555 amortized cost. That's about 370K tokens per day — feasible for a team of developers running continuous agentic workloads, but far beyond what most individual developers or small teams consume.

The privacy argument changes the math entirely for regulated industries. Healthcare organizations, law firms, and defense contractors can't send patient records, case files, or classified data to cloud inference providers — or won't accept the compliance overhead. For them, local inference isn't a cost optimization. It's a deployment requirement. And the Mac Studio is the only desktop-class option that can hold a frontier model.

The Bigger Picture: Why This Matters Beyond One Model

The Muse Spark Open release is a data point. The pattern is the story.

In the past six months, the open-weight frontier has accelerated dramatically. Kimi K3 from Moonshot (2.8T parameters, 50B active), GLM 5.3 from Zhipu, DeepSeek v4.1 Flash, and now Muse Spark Open from Meta. Each model is bigger than the last. Each demands more memory. And with each release, the hardware that can run these models locally without offloading to system RAM gets more exclusive.

NVIDIA's desktop ceiling is the RTX 5090 at 32 GB of GDDR7 — fast bandwidth (1,792 GB/s), but a hard VRAM wall. The DGX Spark pushes to 128 GB but at 273 GB/s — less than a quarter of the M5 Ultra's bandwidth. AMD's Strix Halo sits at 128 GB and 256 GB/s.

Apple's M5 Ultra at 512 GB with 1.2 TB/s is in a class by itself for desktop inference. And Apple knows it. The timing of the MLX distributed-inference paper — published the same day as Meta's model release — is not coincidence. It's coordination.

This is the flywheel: bigger open-weight models drive demand for bigger unified-memory machines. Apple is the only vendor selling those machines at desktop scale. Meta is the most prolific producer of frontier open-weight models. Neither company needs a contract. The market structure is the contract.

ℹ️ Historical parallel: In 1993, Intel and Microsoft didn't have a partnership agreement. They had something better — mutual dependence. Intel needed Windows to sell chips. Microsoft needed Intel to sell Windows. The "Wintel" alliance dominated the PC industry for two decades without a single joint venture. The Meta-Apple inference stack is the same structural pattern, emerging in a different era.

For more context on how we got here, see our earlier pieces: M5 Ultra + Qwen 125B: The Rig That Beats the DGX Spark, Your Next AI Workstation Is a Mac Studio, and Local AI Just Became the Default.

What This Means for You

If you're spending $500+/month on inference APIs and your workload is dominated by a single model family (not routing across specialists), run the amortization math for your token volume. The Mac Studio makes financial sense above roughly 370K output tokens per day. Below that, cloud is cheaper.

If you handle regulated data — healthcare, legal, defense, financial — the Mac Studio is now the only desktop option that can run a frontier model locally without compromise. The compliance benefit alone may justify the hardware cost.

If you're building agentic workflows, watch the Muse Code demos closely. Meta's bet is that the next generation of AI agents will be long-running, tool-using, and locally executed. The one-million-token context window enables workflows that would bankrupt a cloud API budget.

If you're already running local inference on Apple Silicon, the MLX distributed-inference update (requiring macOS 26.2 and MLX 0.30.1+) is worth testing. Single-machine performance on smaller models (45 tok/s on DeepSeek 4.1 Flash, per HN benchmarks) is already practical for many coding workflows. You don't need two machines or Muse Spark Open to benefit from this ecosystem.

If you're invested in NVIDIA hardware, this changes nothing about CUDA's dominance for training or high-throughput serving. But for desktop inference — the "rig on your desk" use case — Apple's memory architecture advantage is now structural, not speculative. The gap will widen as models grow.

The open-weight frontier has found its distribution partner. Meta builds the models. Apple sells the silicon. And for the first time, a nearly-trillion-parameter model runs on hardware that fits under a desk, draws less power than a toaster, and never sends a token to the cloud.

That's not a product announcement. That's a platform shift.

Originally published at ComputeLeap