The photo was deliberately ordinary: Mark Zuckerberg posting a picture of two Mac Studios sitting on a desk, running Meta’s newest AI model at full scale. The ordinary framing was the point. On October 15, 2026, Meta released Muse Spark Open — a native multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active per token, and a context window up to 1 million tokens — and announced that the entire model runs locally on Apple hardware, thanks to a collaboration that would have been unthinkable a year earlier and a new version of Apple’s open-source MLX framework released the same day.
Meta calls it the most capable U.S. open-weight model to date. The more interesting claim is the venue: a near-trillion-parameter, frontier-class, agent-oriented model whose reference deployment is not a hyperscale datacenter but a pair — or small cluster — of machines you can put in an office.
What Muse Spark Open Is
The model is positioned squarely at the workloads that now define the frontier conversation:
- Architecture: sparse mixture-of-experts, 816B total / 42B active, meaning each token routes through a small fraction of the model — the same efficiency logic that makes very large models economical to serve.
- Modality and context: natively multimodal (text and image), with context up to 1 million tokens, built explicitly for coding, long-horizon agent tasks, complex software engineering, and multi-step reasoning.
- Meta’s own benchmark claims: first place among frontier open models in six categories the company published — workflow/agentic tasks, tool calling, knowledge reasoning, knowledge reliability, complex visual reasoning, and image-based factual question answering.
- The caveat Meta itself published: on software development and some other tasks, Chinese open-weight models from Kimi, Qwen, and Zhipu remain ahead. That concession matters; the open-weight frontier is genuinely contested, and the “strongest open model” title now comes with fine print about which benchmark day and which task family.
Weights are downloadable for self-hosting, the model is also available via API through partner platforms, and Meta says it has been optimized not only for Apple silicon but for NVIDIA, AMD, and Intel hardware across mainstream inference engines — a deliberate hedge so the launch reads as ecosystem play rather than single-vendor lock-in.
The Engineering: How a Near-Trillion-Parameter Model Fits on a Desk
Apple’s accompanying engineering research, Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio, is the substantive part of the announcement. Running an 816B model outside a datacenter required solving three problems in sequence.
Memory footprint. In native BF16 precision, Muse Spark Open needs more than 1.6TB of RAM for parameters alone — beyond any single machine. In 4-bit NVFP4 quantization, the footprint drops to roughly 493GB, which fits across two top-spec Mac Studios. The new M5 Ultra Mac Studio supports up to 512GB of unified memory (the 512GB configuration goes on sale in late October; the 256GB version is available now).
Interconnect. Splitting a model across separate computers only works if the link between them behaves like a memory bus. macOS 26.2 introduced a low-latency RDMA driver over Thunderbolt 5, and MLX v0.30.1 builds new communication collectives (jaccl) on top of it. Apple’s reference setup connects Mac Studios in a Thunderbolt mesh — two cables per link to double bandwidth — and launches distributed jobs with a single mlx.launch helper.
Parallelism. Apple benchmarked both schemes. With pipeline parallelism, two Mac Studios each hold half the model’s layers; decode runs at 15.74 tokens per second at 8,192-token context, with minimal communication overhead. With tensor parallelism, each layer is split across machines: two Mac Studios reach roughly 25 tokens/second, and a four-machine cluster reaches 36.7 tokens/second — comfortably above average human reading speed. Time-to-first-token for the full model across two machines lands at 8.4 seconds at 8K context, dropping to 4.8 seconds on four machines at the tensor-parallel configuration.
The hardware economics underneath are the M5 Ultra’s 1.2 TB/s of unified memory bandwidth (up from 819 GB/s on the M3 Ultra) and GPU Neural Accelerators that cut compute-bound time-to-first-token by up to 4.6x on smaller models — all at roughly 300W peak power, a fraction of a rack’s appetite. Decode is memory-bandwidth-bound, prefill is compute-bound, and Apple now has a story for both.
The Odd Alliance and Why Both Sides Need It
The strategic subplot is that Apple and Meta spent years publicly hostile — privacy accusations from one side, invention mockery from the other — and still ended up co-announcing a product. Both sides got something concrete:
- Meta gets a local-deployment story that differentiates open weights. “Runs at full scale on a developer’s desk” is a stronger on-the-ground pitch than another hosted benchmark, and it hands the open-weight ecosystem a reference architecture that does not depend on NVIDIA’s allocations. Executives with ties to both companies — Meta’s compute co-lead is a former Apple director — reportedly bridged the relationship.
- Apple gets a frontier-class proof point for the Mac as an AI development machine. For most of the GPU era, serious local model work meant Linux workstations and datacenter accelerators. A near-trillion-parameter model generating 36 tokens/second on Mac hardware is a direct argument that the Mac is not a client device for cloud AI but a viable AI compute node — for agentic coding in particular, where private local environments and large codebase contexts are the daily workload.
Neither company’s incentives require the alliance to extend further. Meta remains platform-agnostic on inference hardware; Apple remains neutral about which models its silicon runs. It is a coalition of convenience around one shared proposition: frontier inference should not be exclusively a datacenter activity.
The Cost Math — and Who Actually Buys This
Desk-scale frontier is not cheap. A single 256GB Mac Studio configuration already exceeds $10,000; Apple’s own two-machine reference implies two 512GB machines that will not even ship until late October; four 256GB units cluster past $40,000. Against that, Meta and Apple both point to the same arithmetic: teams with sustained agent-development workloads currently rent cloud accelerators continuously, paying not just compute but egress, queueing latency, and the operational tax of round-tripping private code and data to a remote endpoint.
For that audience, local hardware pays back on three axes at once:
- Token economics over time — capital expenditure replacing metered inference for repeated, high-volume workloads.
- Data control — source code, customer data, and internal documents never leave the machine, which is the privacy argument that matters most to regulated enterprises.
- Operational independence — local inference is immune to upstream API incidents, throttling, or price changes, a point that landed with unusual force the very same week as DeepSeek’s multi-hour API outage.
The realistic buyer is not an individual hobbyist. It is the AI engineering team, the financial or medical software shop, the government-adjacent integrator, and the academic lab — exactly the customers for whom a self-hostable agent runtime is infrastructure rather than experimentation.
What Moves, and What Doesn’t
It would be overreading the launch to declare the datacenter dead. Four Mac Studios deliver 36 tokens/second for one model instance; a rack serves thousands of concurrent users with elasticity desks cannot match. Quantization to NVFP4 is a real, if increasingly small, quality trade-off, and distributed multi-node inference is still early-adopter infrastructure, not one-click software. Cloud inference will remain the default for elastic, consumer, and lightweight workloads.
What has genuinely moved is the boundary of what “local” can mean. Until now, local AI meant small models with compromised capability — the 7B and 27B class, useful but explicitly subordinate to frontier APIs. Muse Spark Open plus M5 Ultra plus MLX establishes a new tier: frontier-adjacent, agent-capable, million-token-context models owned and operated by their users, on hardware that sits on a desk and draws 300 watts. Add the same-day open-weight release of other large agentic MoEs and the arc becomes clear: the open-weight world is compressing the capability gap at the top while the hardware world expands what fits at the edge.
That convergence is the DeepThink thesis as well — that efficient, transparent, portable reasoning is the load-bearing capability of the agent era, and that its value grows as reasoning escapes the datacenter. When a near-trillion-parameter model can run in the room next to you, the question stops being whether local frontier is possible and becomes how the whole industry reorganizes around the fact that it is.