Can you run Muse Spark Open on a Mac?

On October 15, 2026, Meta released Muse Spark Open, an open-weight mixture-of-experts model with 816B parameters, of which about 42B work on each token. It has 62 layers and a 1,048,576-token context window, and it is natively multimodal.

It is not the Muse Spark you chat with on meta.ai, which stays closed. The open model "shares the same base model as Muse Spark 1.3" (PyTorch).

• Licence: the Muse Model License. Open weights, not open source. Products built on it must show "Built with Muse Spark Open". Offering it to others as a model service needs a separate agreement with Meta once your company's revenue passes $20M a year. Gated download. Meta approves each download request by hand on Hugging Face.
• Meta's own builds: BF16 (1,644.3 GB) and NVFP4 (533.0 GB). The MLX builds come from mlx-community; the GGUF builds for llama.cpp come from ggml-org and unsloth.
• Ollama lists Muse Spark Open only as a cloud model, run on Ollama's servers and billed per use, not as a local download.

Mark Zuckerberg posted on launch day: "Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple." John Ternus, Apple's CEO (Apple leadership), replied the same day. Apple's side of the launch is a research post that measures the model on M5 Ultra Mac Studios.

The limit is what macOS lets the GPU use, not the RAM on the box: about 239.1 GB on a 256 GB Mac and 478.3 GB on a 512 GB Mac by default (macOS 26 and later). The 256 GB figure rests on weak evidence and the 512 GB one is assumed, since no log confirms it yet.

sudo sysctl iogpu.wired_limit_mb=459122

Context memory is small for a model this size. 46 of the 62 layers only look back 2,048 tokens, so only 16 layers keep the whole conversation, 65,536 bytes per token at 16‑bit. That comes to 0.9 GB at 8,192 tokens, 4.7 GB at 65,536 and 69.1 GB at the full 1,048,576. mlx-lm caps the windowed layers this way; estimates of hundreds of GB at full context ignore the window.

The 4‑bit build does not fit one Mac, so its only speeds are Apple's, with the model split across M5 Ultra Mac Studios: 16 to 25 tokens per second on two Macs and 37 on four, at 8,192 tokens of context.

On a single Mac, the only published speeds come from Hugging Face, a launch partner, on a 512 GB M3 Ultra: 18 tokens per second with ggml-org's 2‑bit Q2_K build in llama.cpp (304 GB used by the model), and 21 with the 3‑bit oQ3e build in oMLX, both at short context. We found no independent repeat of Apple's or Hugging Face's numbers.

We do not estimate Muse speeds. Our memory-bandwidth formula predicts 25 tokens per second for the 2‑bit run Hugging Face measured at 18, about 40% too fast, so for mixtures of experts above 120B parameters this page shows measured speeds or "Not measured".

On one 512 GB Mac, every build from Q2_K (311.1 GB) to MLX 3‑bit (445.7 GB) fits under the 478.3 GB default. Speeds exist only for the M3 Ultra (18 and 21 tokens per second, above); no single M5 Ultra figure has been published.

On one 256 GB Mac (any M5 Ultra, on sale today), only UD-IQ2_XXS (231.0 GB, 97% of the 239.1 GB default) and the two 1‑bit builds fit, with the macOS 26 and later default; a 256 GB M3 Ultra still on macOS 15 or earlier gives the GPU only 206.2 GB, and the 2‑bit build no longer fits. Hugging Face notes that "even the 3‑bit quantization requires more than 256 GB of RAM". Nobody has published a speed or quality figure for these builds.

Commands for one 512 GB Mac, from Hugging Face's launch post. llama.cpp with ggml-org's 2‑bit build (a 304.1 GB download):

curl -LsSf https://llama.app/install.sh | sh

llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K

Or mlx-lm with the 3‑bit oQ3e build (371.5 GB). On one Mac, mlx_lm.generate is fine. Hugging Face's 21 tokens per second for this build was measured with oMLX, not mlx-lm.

mlx_lm.generate --model mlx-community/Muse-Spark-Open-816B-A42B-oQ3e --prompt "Explain convolutional networks like I am 5" --max-tokens 5000

Choosing between the Mac Studios: which Mac Studio to buy for AI.

Muse Glimmer 30B is Meta's smaller open model, released in August 2026 under Apache 2.0 with no download gate. It is dense, not a mixture of experts: all 29.78B parameters (1.8B of them a vision encoder that ships as a separate file) work on every token. It has 52 layers, 39 of them windowed, a 131,072-token context, and reads text and images.

Meta's deploy guide says its Q4_K_M file "measured about 24.1 GiB at one request and 16K context, so leave room for the OS and Bionic". That is 25.9 GB, 8.1 GB more than our formula gives for the same file and context (17.8 GB), and Meta does not say whether the vision file was loaded. We grade Meta's file on Meta's figure: it does not fit a 24 GB Mac (19.1 GB for the GPU), takes 96% of the 26.8 GB a 32 GB Mac allows, and leaves room on 36 GB (30.2 GB). SGLang's team, in the same guide, verified Muse Glimmer on an M5 Pro with 64 GB and lists 48 GB as the minimum for its MLX builds; the 32 GB figure here rests on Meta's LM Studio measurement, not on a verified setup.

• 24 GB: run text only in llama.cpp, with --no-mmproj (which skips the vision file) and a short context. By our formula, not measured, Meta's 17 GB file then needs about 17.7 GB, 93% of the 19.1 GB the GPU can use, and unsloth's UD-Q4_K_XL file (15.9 GB) about 16.8 GB (88%), a little more headroom. Meta's file with its vision file and DFlash draft model is 19.8 GB of files, more than a 24 GB Mac gives the GPU.
• Meta says its 17 GB quantization targets 24 GB of VRAM. That means a graphics card with 24 GB to itself; a 24 GB Mac shares its memory with macOS and lets the GPU use 19.1 GB.
• 16 GB: only the 2‑bit build fits (11.6 GB of the 12.7 GB the GPU can use), and nobody has published how much quality it loses.
• Speed: a dense 30B model is slow on base chips. Our estimate for Meta's file is 7.6 to 8.9 tokens per second on a 32 GB M6 and 26 to 31 on a 48 GB M5 Max with the 40-core GPU. Meta measured 23.7 on an M4 Max and 26.6 on an M5 Max, faster still with its DFlash draft model, but with ExecuTorch rather than llama.cpp, LM Studio or Ollama, and without naming the GPU.

To run it with Ollama 0.32.7 or later, use Ollama's MLX build for Apple silicon (4‑bit NVFP4 with the DFlash draft model, a 19 GB download), or the standard 18 GB build. Both include the vision model, so on a 24 GB Mac use llama.cpp text only, below. Ollama serves an API at http://localhost:11434.

ollama run muse-glimmer:30b-mlx

ollama run muse-glimmer:30b

In LM Studio, download Muse Glimmer from the model search, then start the local server on port 1234:

lms server start

llama.cpp needs build b10353 or newer. Meta's command uses a 131,072-token context; keep it shorter on 32 GB, where Meta's figure at full context would not fit.

llama serve -hf meta-models/Muse-Glimmer-30B-GGUF --hf-file Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf -ngl 99 -c 32768

llama serve -hf unsloth/Muse-Glimmer-30B-GGUF:UD-Q4_K_XL --no-mmproj -ngl 99 -c 8192

This is the setup from Meta's Mac cookbook, with the RDMA steps from the MLX docs and Apple's technote, and one fix to the last step.

1. Get the Macs: Two M5 Ultra Mac Studios with 512 GB each (not on sale until Apple ships the option) or four with 256 GB each, two Thunderbolt 5 cables between every pair of Macs (Apple's measured setup; one cable works but is slower), and room for the 533.0 GB download on each Mac.
2. Update macOS: Install macOS 26.2 or later on every Mac. RDMA over Thunderbolt, the fast link MLX uses, starts there.
3. Turn on RDMA: On each Mac, start up in macOS Recovery, choose Utilities, Terminal, run rdma_ctl enable and restart. Check it with ibv_devices. This cannot be done remotely.
4. Turn off the Thunderbolt bridge: In System Settings, Network, make Thunderbolt Bridge inactive on every Mac, as Apple advises for Macs cabled in loops.
5. Set up SSH: From the first Mac, set up passwordless SSH to every other Mac with ssh-keygen and ssh-copy-id.
6. Request the weights: Ask for access to Meta's NVFP4 repo on Hugging Face (Meta approves each request by hand), then log in with your Hugging Face access token on every Mac.
7. Install mlx-lm: Install mlx-lm 0.32.2 or later on every Mac. It is the first release that loads the public checkpoint, and it brings MLX 0.32.2 or later with it.
8. Write the host file: On the first Mac, run mlx.distributed_config over Thunderbolt with the jaccl backend to write hostfile.json.
9. Start the model: Run mlx.launch with mlx_lm.server for an OpenAI-compatible API on port 8080, or with mlx_lm.chat to chat in Terminal. Never with mlx_lm.generate.

rdma_ctl enable

ibv_devices

ssh-keygen -t ed25519
ssh-copy-id mac-studio-2.local
ssh mac-studio-2.local 'echo ok'

pip install -U "mlx-lm>=0.32.2"

mlx.distributed_config --verbose --hosts mac-studio-1,mac-studio-2 --over thunderbolt --backend jaccl --auto-setup --output hostfile.json

mlx.launch --hostfile hostfile.json --backend jaccl -- mlx_lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080

mlx.launch --hostfile hostfile.json --backend jaccl -- mlx_lm.chat --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4

The model reasons at its highest level by default, so leave plenty of room for output tokens, as Meta's cookbook advises.

No. PocketWebTools for Mac runs 6 fixed chat models on llama.cpp, the largest Qwen3.6 35B‑A3B (a 22.4 GB download). It cannot load your own model files, run MLX or link Macs, and it does not ship Muse Spark Open or Muse Glimmer. Use the free tools above for those.

What it does instead: private chat with its own models (Qwen3.6 35B‑A3B measured at 25 to 27 tokens per second on a 32 GB Mac mini M4), transcription with speaker labels, text to speech and voice cloning, OCR and photo upscaling, all offline. See what the app does.

If you have two 512 GB or four 256 GB M5 Ultra Mac Studios, run Meta's 4‑bit build with mlx-lm 0.32.2 or later and expect Apple's measured 16 to 25 tokens per second on two Macs and 37 on four. One 512 GB Mac runs the 2‑bit build at about the 18 tokens per second Hugging Face measured; a 256 GB Mac is an untimed squeeze.

On any other Mac, run Muse Glimmer 30B with Ollama, LM Studio or llama.cpp: 32 GB at the least by Meta's LM Studio figure, 36 GB or more for room to spare, and 48 GB for the setup SGLang verified. A 24 GB Mac can run it text only in llama.cpp with --no-mmproj (our estimate).