Facts you can check on 16 October
- On 15 October 2026 Meta released the weights for Muse Spark Open: 816 billion parameters in total, 42 billion active per token, and a 1 million token context at inference. Input is text, image, video, and audio. Output is text. Post: Meta AI Research. Card: Hugging Face.
- The Mac path is not one computer. The documented setups are two M5 Ultra Mac Studios with 512GB of unified memory each, or four with 256GB each. The checkpoint is NVFP4, the runtime is MLX 0.30.1+, the OS is macOS 26.2+, and the machines talk over RDMA on Thunderbolt 5. Table: Hardware platforms.
- Apple’s MLX post the same day has the speeds. At a context of 8,192 tokens, four machines in tensor parallel: 4.8 seconds to the first token, then 36.73 tokens/second. That is
mlx_lm.benchmark, batch 1, 128 generated tokens. It is not a kvmboot run, and it is not the 1 million token window. Post: Apple Machine Learning Research. - Apple’s 25 August newsroom post says a Mac Studio with 512GB of unified memory arrives in late October. M5 Ultra starts at $5,499. That is the base config, not the 256GB or 512GB price. Post: Mac Studio newsroom.
- The weights ship under the Muse Model License. Many uses are free, and the product has to show “Built with Muse Spark Open”. If you sell the model as an API and the company plus affiliates gross more than $20 million in any 12 months, you notify Meta and sign a separate agreement first.
816 billion is the total. The Mac run is 4-bit NVFP4, and each token activates about 42 billion parameters. One Mac Studio does not hold it. A Mac mini does not come close.
15 October shipped Muse Spark Open, not a Mac that holds 816 billion parameters
Meta’s research post calls Muse Spark Open the most capable US open-weight model to date, built for coding and long-horizon agents. The developer post the same day describes a sparse mixture-of-experts: 816 billion parameters in total, about 42 billion active per token, input across text, image, video, and audio, and a 1 million token context. Weights went to Hugging Face the same day, and also to Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry. APIs were listed on day one at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Developer post: Build with Muse Spark Open.
The model card fixes the shape: a MoE transformer plus a perception encoder, vocabulary 202,048, training data from more than 100 languages. Of 256 experts, each token uses 8. Forty-two billion over 816 billion is not 8 divided by 256, because the attention layers are shared. Native precision is BF16. The checkpoint Apple timed on Mac is NVFP4. Recommended sampling is temperature 1.0, top_p 1.0, top_k off. Coding and agents are told to use high or xhigh, and xhigh is the default. Those sentences are on the card. We did not run them.
| Line on the card | The number | What it does to the Mac headline |
|---|---|---|
| Total parameters | 816B | The number in the headline |
| Active parameters | 42B; 256 experts, 8 used per token | Decode mostly reads this slice |
| Context | 1M tokens at inference | Apple’s speed table stops at 8,192 and 65,536 |
| Modalities | Text, image, video, audio in; text out | Not a small text-only model |
| License | Muse Model License | Open weights, not Apache |
So 15 October was two releases at once: the weights, and a path that splits the NVFP4 checkpoint across Mac Studios. Meta says Muse Code, on that desktop setup, shipped iOS changes to Ice Cubes, an open-source Mastodon client, and that the model walked DaVinci Resolve frame by frame. That is Meta’s description. This article does not reproduce it.
How 816 billion parameters land on a Mac: two 512GB machines, or four with 256GB
Apple’s limit is memory. In native BF16 the parameters alone need more than 1.6TB. In NVFP4 the weights are about 493GB. Add the activations and one Mac Studio still cannot hold the model, even with 512GB of unified memory. MLX splits it. Pipeline parallel cuts the model by depth across two machines. Tensor parallel cuts every layer by width. The link is RDMA over Thunderbolt 5, which macOS 26.2 added. MLX 0.30.1 calls that collective jaccl.
Meta’s hardware page and the Mac cookbook name the same machines: two M5 Ultra Studios at 512GB, or four at 256GB. The cookbook budgets about four hours, asks for the full NVFP4 checkpoint on the disk of every Mac, and wants passwordless SSH. Every pair needs a direct Thunderbolt 5 link, and the doc says to use two cables per pair so the link is twice as wide. With four Macs, each machine connects to the other three. Counted from that rule, two machines need 2 cables and four need 6 pairs times 2, which is 12. The cookbook does not print the number 12. Cookbook: Muse Spark Open on Mac.
| Path | Machines | Precision and runtime |
|---|---|---|
| One Mac Studio | Does not hold Muse Spark Open | Apple: BF16 parameters alone exceed 1.6TB; NVFP4 weights are about 493GB, and activations still do not fit one machine |
| The desktop Meta describes | 2× M5 Ultra, 512GB | NVFP4, MLX, about 1TB of unified memory combined |
| The other documented path | 4× M5 Ultra, 256GB | The same 1TB, without waiting on the 512GB option |
| NVIDIA | B300 / GB300: 4 GPUs for NVFP4, 8 for BF16 | SGLang, vLLM, TensorRT-LLM |
| Model-card VRAM floor | One GPU | At least 700GB for NVFP4; at least 2TB for BF16 |
493GB and 700GB are not a contradiction. Apple’s 493GB is about how big the NVFP4 weights are. The Hugging Face card’s 700GB and 2TB are the VRAM they tell you to have on a GPU box, runtime included. The Mac path is NVFP4 only. BF16, on the card, points at GPUs, not at a Mac.
36 tokens/second is four machines at 8,192 tokens, and it is Apple’s table
Apple ran mlx_lm.benchmark on three layouts, context 8,192, batch 1, 128 generated tokens. Muse Spark Open has no single-machine row, because it does not fit. The table below copies only the 8,192 column that the page shows by default. The numbers are Apple’s. They are not a test we ran.
| M5 Ultra layout | Time to first token (seconds) | Decode (tokens/second) |
|---|---|---|
| Two machines, pipeline | 8.4 | 15.74 |
| Two machines, tensor parallel | 8.2 | 24.99 |
| Four machines, tensor parallel | 4.8 | 36.73 |
“Over 36 tokens per second” is the last row, and the figure caption says the context is 8,192. At 65,536 tokens Apple’s prose states a different fact: on four M5 Ultras, two cables per link cut time to first token from about 50.5 seconds to 40.6 seconds. They say tensor parallel still keeps a decode speedup at long context, and they do not print a 65,536 decode number next to 36.73 in that paragraph. The 1 million token window is not in the table.
One multiplier gets pasted onto the wrong model. On a single machine, time to first token for Qwen 3.8 (27B, dense, BF16) and Mistral Medium 3.5 (128B, dense, FP8) is up to about 4.6× on M5 Ultra versus M3 Ultra. Decode is about 1.6×, in line with memory bandwidth moving from 819GB/s to 1.2TB/s. At a context of 65,536 those two models use about 68GB and 158GB on one 512GB machine. Apple puts peak power around 200W for M3 Ultra and 300W for M5 Ultra. None of those sentences are about Muse Spark Open. It has no M3 row and no single-machine row.
Apple wants a desktop cluster it can sell, not 816 billion parameters on an iPhone
The 25 August Mac Studio post already said this: M5 Ultra scales to a 36-core CPU, an 80-core GPU, 512GB of unified memory, and 1.2TB/s of bandwidth. Thunderbolt 5 plus RDMA pools memory across machines, and four systems are described as up to about 3× the inference of one. The 512GB option is “late October”. Other configs start arriving 22 September. The $5,499 starting price ($5,099 education) is the base M5 Ultra, not two 512GB machines.
The 15 October MLX post gives that hardware a model name. Muse Spark Open does not fit one box, so “buy one Mac” is not the story. The story is the cluster. MLX is Apple’s open-source array framework, the Neural Accelerators do the matrix multiplies, and jaccl rides Thunderbolt. The post ends by serving an OpenAI-compatible API on port 8080 and pointing OpenCode at it. Apple wants the desktop demo of a frontier open-weight model to happen on Mac Studio, not only on rented GPUs.
It does not happen on an iPhone, and it is not an Apple Intelligence feature. The 4.6× figure measures time to first token for smaller dense models on one machine. It is not a speedup for Muse Spark Open. Buying a Mac mini because of the headline buys a different unified-memory tier, orders of magnitude under a 493GB checkpoint.
Meta wants the US open-weight name, and a license that pulls large resellers back
“Most capable US open-weight model” is Meta’s sentence. The comparison table on the model card is not a sweep of every row. The rows below are the ones that answer where it leads. The setting is the xhigh effort Meta lists. The scores are the card’s scores. We did not rerun them, and we did not refresh Artificial Analysis on 16 October.
| Benchmark on the card | Muse Spark Open | Higher or nearby scores printed on the same card |
|---|---|---|
| MCP-Atlas | 89.2 | Highest on that row; Qwen3.8 84.5, GLM 5.3 84.2, Kimi K3 82.3 |
| Artificial Analysis Intelligence Index | 39 | Qwen3.8 40, Kimi K3 44, GLM 5.3 45 |
| OSWorld Verified | 80.4 | Kimi K3 84.8 |
| Terminal-Bench 2.1 | 76.4 | Kimi K3 85.0, GLM 5.3 83.9, Qwen3.8 82.0 |
| DeepSWE 1.1 | 60 | Kimi K3 68, GLM 5.3 67 |
The row it leads is tool use, MCP-Atlas. On the intelligence index, computer use, terminal, and software-engineering rows, the card prints a higher number for Kimi, GLM, or Qwen. Meta is aiming the model at agents and long jobs, not at winning every board. The desktop demo supports “you can run it at home”. The seven APIs listed the same day are the door most people can actually open this week.
The license is on the Hugging Face model page. It allows free use, fine-tuning, and redistribution, with two gates. First, any product that uses the software or a derivative has to show “Built with Muse Spark Open” in the interface, the docs, the site, or the settings. Second, if you or an affiliate give a third party model outputs through an API or something like it, that is Model as a Service. If gross revenue of the company and affiliates exceeds $20 million in any consecutive 12 months, you notify Meta and sign a separate agreement before offering it that way. Internal use, research, and using the model as a feature inside your own product — not selling the model itself as the service — skip that second gate. The usage policy says it is not intended for anyone under 18. Open weights with a commercial gate, not Apache under a new filename.
If you are not buying that Mac Studio cluster, three paths still match the docs
The headline asks what the two companies want. On your own hardware, only the lines below match the documents as of 16 October.
- You want to run Muse Spark Open itself: two 512GB M5 Ultras or four 256GB ones, macOS 26.2, MLX 0.30.1, a direct Thunderbolt 5 link on every pair. The 512GB option is still the “late October” line from the 25 August post. Four 256GB machines do not wait on it.
- You want it this week and you are not buying a cluster: use one of the seven APIs named in the launch post. That post has no price list, so this article does not invent a rate.
- You already have one 512GB Mac Studio: on Apple’s table, Qwen 3.8 and Mistral Medium 3.5 fit. Muse Spark Open does not.
- You have a Mac mini: its unified memory is an order of magnitude under a 493GB NVFP4 checkpoint. It can be the Mac that calls an API, compiles, and signs. It cannot be the machine that holds the weights. Why memory runs out first is in the related piece at the end.
Two things are still open: whether the 512GB Mac Studio actually goes on sale in late October, as the August post said, and what the hosted APIs bill. Apple also has not printed, in that paragraph, a 65,536-token decode speed in the same column as 36.73 tokens/second. Until that number exists, do not describe the 8,192 result as the speed of the 1 million token window.
Questions
Can one Mac mini run Muse Spark Open?
No. Apple puts the NVFP4 weights at about 493GB, and with activations even one 512GB Mac Studio does not hold the model. The documented Mac path is two 512GB M5 Ultra Mac Studios, or four with 256GB.
Does inference compute all 816 billion parameters every token?
No. It is a mixture of experts. The card says 256 experts and 8 active per token, about 42 billion active parameters. 816 billion is the total.
What are the conditions on 36 tokens/second?
Apple’s table: four M5 Ultras, tensor parallel, NVFP4, context 8,192, mlx_lm.benchmark, batch 1, 128 generated tokens. Not the 1 million token window, and not a test run for this article.
Can you use the weights commercially?
The Muse Model License allows many free uses, and the product must say Built with Muse Spark Open. Selling the model as an API to third parties, once the company and affiliates gross over $20 million in any 12 months, requires notice to Meta and a separate agreement. Internal use, and embedding it as a feature of your own product, skip that clause.
Is the 4.6× figure Muse Spark Open versus the previous Mac Studio?
No. 4.6× is time to first token for Qwen 3.8 and Mistral Medium 3.5 on one machine, M5 Ultra versus M3 Ultra. Muse Spark Open has no single-machine row and no M3 row.
The cloud Mac sits on the side that calls the model
The NVFP4 weights for Muse Spark Open do not fit a cloud Mac mini. Unified memory at 16GB or 24GB is fixed when you order. That pool is for Xcode, the simulator, and an agent that calls OpenRouter or the MLX server you stood up somewhere else. The machine is Apple silicon running macOS, so compile and sign stay on the native toolchain, and the idle draw is low enough to leave it on. What you change is the memory tier. You do not start by buying two 512GB Mac Studios that the August post still had arriving in late October.