What if your AI didn't need the cloud?

What was actually released

Muse Spark Open is an open-weight model with 816 billion total parameters, of which 42 billion are active for any given token. It is a mixture-of-experts transformer - 256 experts, 8 of them used per token - with a 1 million token context window, and it accepts text, images, video and audio as input.

Meta says it is optimised for coding and agentic work: long-horizon tasks, software engineering, multi-step reasoning.

Meta also describes it as the most capable US open-weight model to date. That is a claim about a field, made by the company that released the model, and we will come back to it.

One small thing worth noting, because this is the sort of detail that tends to matter later: the model card says 816 billion parameters, while the Safetensors metadata on the same page lists 822 billion. We do not know which is right.

The Mac Studio demonstration, and what it does not mean

The part that caught everyone's attention is that a model of this size ran on desktop computers.

Meta's announcement says Muse Spark Open runs on two Mac Studios with M5 Ultra chips using Apple's MLX framework, and that it is the first model in its class to run on a desktop this way. Meta demonstrated agentic work on that setup - changes to an iOS app called Ice Cubes, and video cropping in DaVinci Resolve.

Apple separately published its own measurements of distributed inference on Mac Studio hardware. Those are worth reading alongside Meta's, because they are more specific and in one respect less flattering.

Apple's headline throughput figure - decode above 36 tokens per second - comes from four Mac Studios, not two. Apple tested one, two and four machines, linked by RDMA over Thunderbolt 5 using a low-latency driver introduced in macOS 26.2. Peak power on the M5 Ultra was around 300 watts per machine.

Memory is the real constraint, and it is where the widely repeated "1TB" figure gets slippery. Each Mac Studio holds up to 512GB of unified memory, so two of them do give you a terabyte. But the model needs more than 2TB of VRAM at full BF16 precision. It only fits at all because it is quantised: Meta's card asks for at least 700GB in NVFP4, and Apple puts the NVFP4 weights at around 493GB.

So the honest version is not "an 816 billion parameter model runs on two desktops". It is "a compressed version of it does, slowly, across machines costing tens of thousands of pounds, and Apple's own best numbers used twice as many".

That is still a real achievement. It is just a narrower one than the headlines.

Is Apple actually involved?

This is worth pausing on, because almost every write-up has described it as a Meta-Apple collaboration.

Meta credits Apple's MLX framework, and says the model was optimised for it. Apple's own research page, which benchmarks the model in detail, never mentions Meta in its prose at all - the word appears only inside the file path of the model it downloaded.

TechCrunch's summary says the two companies "teamed up", but its own article describes no formal partnership.

What is evidenced is narrower than a partnership: Meta optimised for Apple's framework, and Apple benchmarked a publicly released model on its hardware. That may well have involved cooperation. But "Meta and Apple have teamed up" is currently a characterisation rather than something either company has stated.

The benchmark picture is mixed, and it comes from Meta

Meta's own comparison table puts Muse Spark Open against three Chinese open-weight models: GLM 5.3, Kimi K3 and Qwen 3.8.

It does not win.

It leads on JobBench (61.8 against 61.4, 54.3 and 52.7), on MCP-Atlas (89.2 against 84.2, 82.3 and 84.5) and on Humanity's Last Exam without tools (26.1 against 16.4, 22.2 and 17.3).

It trails on the Artificial Analysis Intelligence Index (39 against 45, 44 and 40), on DeepSWE (60 against Kimi's 68 and GLM's 67) and on Terminal-Bench 2.1 (76.4 against 85.0, 83.9 and 82.0).

Three caveats apply, and Meta states them itself. The comparisons use different reasoning settings across models. Some cells are missing for some competitors. And the Qwen entry is a considerably larger model - 2.4 trillion total parameters with 95 billion active - so it is not a like-for-like row.

One result stands apart. On StrongREJECT v2, an adversarial safety test where lower is better, Muse Spark Open scores 0.5 against 62.9 to 66.6 for the other three. That is not a narrow margin, it is a different order of magnitude, and it is the single most interesting number in the table. It is also Meta's own measurement of its own model's safety post-training, published without an independent evaluation alongside it. Meta says a full evaluation report is coming.

Benchmark tables are the easiest thing in AI to over-read - how capability is actually measured is worth the ten minutes before trusting any of them.

Open-weight is not open-source, and here the difference is specific

Meta released the weights - the learned numerical settings that make the model work. Developers can download them, run the model on their own hardware and fine-tune it.

That is not the same as open-source, and in this case the gap has a number attached to it.

The licence is the Muse Model License, dated 15 October 2026. It permits use, modification and distribution, but derivatives must carry the words "Built with Muse Spark Open". And if you offer the model as a service and your aggregate gross revenue passes 20 million US dollars in any twelve consecutive months, you need a separate written agreement with Meta.

For almost everyone that threshold is irrelevant. For anyone building a business on it, it is the whole point: this is a commercial licence with a free tier, not a public good. The training data and the full development process are not published either.

Why running AI locally matters anyway

Strip away the overstatement and something real remains.

When you use a cloud AI service, your question travels to someone else's computers, and you depend on their infrastructure, their availability, their pricing and their data handling. For most questions that is a perfectly good trade. For a solicitor working through privileged documents, or a clinic handling patient records, or a business whose contracts cannot leave the building, it is a harder one.

A model that runs on hardware you own changes that calculation - though not as cleanly as it sounds.

Local processing does not automatically mean private. An application built around a local model can still send data out through its tools, integrations and telemetry. What local control gives you is the option, not the guarantee.

Nor is it cheaper by default. Two or four Mac Studios, the electricity to run them, and someone to maintain them is a real cost, set against an API bill that starts at nothing.

What it does offer is independence from a single supplier, and the ability to build something shaped around one organisation rather than around a general-purpose chatbot.

The part that deserves more attention than the hardware

Muse Spark Open is built for agentic work - models that do things rather than answer questions. Examine files, write code, update records, work through a sequence of steps. If you are unclear on the distinction, what an AI agent is covers it plainly.

Running an agent locally gives you more control over the environment it operates in. It also removes something.

When a model is served from a provider's infrastructure, that provider can update its safeguards, withdraw access or switch off a capability. Once the weights are on your drive, nobody can do any of that. The safety post-training that produced that 0.5 score is baked in at release and cannot be revised afterwards by the people who built it.

That cuts both ways, and the direction depends entirely on who is holding the drive.

It also makes the question of permissions sharper rather than softer. An agent with access to your files and your applications can cause real damage through ordinary error, and there is no vendor to ring.

Controlling where an AI runs is not the same as controlling what it does.

Does this bring AGI closer?

Not in the sense people usually mean.

Nothing here is evidence of a new level of general capability. Running a very large model on desktop hardware is an advance in deployment and efficiency - in compression, memory bandwidth and interconnect speed - rather than in intelligence. The model's own benchmark table has it losing to three competitors on three of six measures.

But capability spreading is its own kind of progress. A system that needed a data centre last year and needs two desktops this year will need something smaller next year. What changes is not how clever the thing is, but how many people can run one, adapt one, and build on one without asking permission.

That is the thread connecting this to the week's other big story. China has just proposed export controls on frontier AI that treat a set of model weights as a controlled item - something that can be transferred, licensed and restricted. Meta's release is the same object viewed from the other end: weights as something you publish, and anyone with the hardware can hold.

Both stories are about the same question. Not how intelligent AI becomes, but who ends up holding it.

Common questions

Can I run this on a MacBook?

No. The full-precision model needs more than 2TB of VRAM, and the compressed version needs roughly 493GB to 700GB. Meta's configuration is two Mac Studios with M5 Ultra chips; Apple's best throughput figures used four.

Does running AI locally make my data private?

Not on its own. It removes the need to send every request to a model provider, but the application around the model can still transmit data through its tools, integrations and telemetry.

Is it free?

The weights are free to download and the licence covers most developers at no cost. If you offer it as a service and pass 20 million US dollars of revenue in twelve months, you need a separate agreement with Meta. The hardware and electricity are not free either.

Is it the best open-weight model available?

Meta says it is the most capable US open-weight model. On Meta's own table it trails GLM 5.3, Kimi K3 and Qwen 3.8 on three of six capability measures, so "best available" is not supported - including by Meta, which worded the claim carefully.

Will cloud AI become unnecessary?

No. Cloud services offer scale, convenience and hardware almost nobody can run at home. Local and cloud AI will coexist; the significance here is that there is now a genuine choice at the top end of capability.

Sources

• Meta: Introducing Muse Spark Open, the most capable US open-weight model to date (15/10/26)
• Apple Machine Learning Research: Local and distributed large LLM inference with MLX and RDMA on Mac Studio
• Hugging Face: meta-models/Muse-Spark-Open-816B-A42B model card and licence
• TechCrunch: Meta launches an open-weight AI model that runs on Macs (15/10/26)