Muse Spark Open on a Mac: What Runs on One 512 GB Mac Studio

What Meta shipped

Meta released Muse Spark Open on 15 October 2026, with the weights on Hugging Face behind a manual-approval form and the announcement on its research blog. The ready-made GGUF and MLX quantizations from ggml-org, Unsloth and mlx-community are ungated and carry the same licence.

The card’s benchmark table is vendor-reported, at xhigh effort. Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, and Artificial Analysis lists the same 39 on its own page (10th of 120 models when checked on 18 October). Meta calls it the most capable US open-weight model to date. That is a claim about US models: on the card’s own table, GLM 5.3 (45), Kimi K3 (44) and Qwen3.8-2.4T-A95B (40) score higher.

It leads that table on MCP-Atlas (89.2), JobBench (61.8) and HLE-Diamond without tools (26.1). Coding numbers need their harness attached. Terminal-Bench 2.1 is 76.4 with the Terminus 2 agent. DeepSWE 1.1 (60) and SWE-Atlas-QnA (58) ran on each vendor’s “preferred harness”, meaning each lab’s own scaffold. The card gives no SWE-bench Verified or SWE-Bench Pro score.

In the LLMCheck leaderboard, capability is near the ceiling at 49 of 50. The overall score is modest: LLMCheck Score 65, #42 of 111. The other three parts of the formula penalise an 816B model: speed 8 of 25, from the estimate on the M5 Ultra (the reference chip for models that need more than 128 GB); memory 3 of 15, for a 342 GB minimum; and licence 5 of 10.

The Muse Model License in plain English

The licence text, dated 15 October 2026, starts from MIT-style wording: use, copy, modify, fine-tune, distill, merge, redistribute and sell. Four conditions follow:

• A credit. Every product or service that uses the model, or a Derivative, must prominently show “Built with Muse Spark Open”. The credit can go in the UI, docs, website or settings. A Derivative includes any model trained or distilled on its outputs.
• A revenue gate for API hosts. A company offering models as a service whose group revenue tops US$20M in any 12 months must notify Meta and sign a separate agreement first. Internal use, research, non-API commercial use and using it inside your own non-API product are exempt.
• A usage policy. Exhibit A lists prohibited uses and says the model is not intended for people under 18. It ships with every copy.
• Outputs are yours. Meta claims no ownership of what the model generates.

For someone running it on their own Mac, nothing here costs money. The credit only matters if you ship something built on it. The index scores the licence 5 of 10, level with Meta’s Llama licence and the Qwen and Moonshot custom licences (methodology).

What one 512 GB Mac runs today

Two single-Mac figures have been published, both in Hugging Face’s launch post of 15 October and both on a Mac Studio M3 Ultra with 512 GB. That is the only Mac that can host the model today. Apple says the 512 GB M5 Ultra configuration is coming in late October (9to5Mac).

Muse Spark Open on one Mac Studio. Published single-Mac figures plus the index’s M5 Ultra estimate.

oMLX’s 0.7.1rc1 release notes report the same oQ3e result: 21.3 tok/s decode and 225 tok/s prefill on the M3 Ultra (published).

One detail matters if you download today. mlx-community re-uploaded the oQ3e build on 16 October in mlx-lm 0.32.2’s layout, keeping the output head in BF16. Its card puts the current files at 3.62 bits per weight on average and 371.5 GB on disk, and says they take about 372 GB in oMLX. That is about 2 GB more than the original upload Hugging Face timed (369.5 GB on disk; the blog’s “about 344 GB” is the same build in GiB), and no speed has been published for the re-upload yet.

The index labels both runs sourced: a published figure from an accountable source that documents its setup. The 512 GB M3 Ultra page carries sourced18 tok/s as its figure for that Mac. The leaderboard quotes the M5 Ultra estimate instead, because that is its reference chip for models that need more than 128 GB. Apple’s own single-Mac tables back the scaling up. In the same week Apple published Qwen3.8-27B (BF16, 8,192-token prompt) decoding at 19.69 tok/s on an M5 Ultra against 12.46 tok/s on an M3 Ultra, about 1.6 times faster (both published).

Memory headroom decides how much context you get. The Q2_K build’s 304 GB leaves roughly 200 GB of a 512 GB Mac for macOS and the KV cache. The re-uploaded oQ3e leaves about 140 GB. Neither source publishes a long-context speed, so read the figures above as short-prompt numbers.

Why 4-bit needs two Macs

Meta’s model card says the model is optimized to run on two M5 Ultra Mac Studios using MLX. Meta’s blog describes that pair as one machine with a terabyte of unified memory. Apple’s machine-learning team explains the reason in its 15 October post on distributed MLX inference. At NVFP4 the parameters alone take about 493 GB, and once working memory is added the model no longer fits on one 512 GB Mac.

Apple published decode speeds for clusters of 512 GB M5 Ultra Mac Studios. The Macs were linked over Thunderbolt 5 with RDMA through MLX’s jaccl backend, which needs macOS 26.2 or later and MLX 0.30.1 or later. Each figure is mlx_lm.benchmark at an 8,192-token prompt with 128 generated tokens:

Multi-Mac cluster figures, published by Apple (NVFP4). They are not single-Mac results, and the index does not store them as per-chip rows.

Two cautions before setting these beside the single-Mac table. The workloads differ: an 8K-token prompt here, 1K there. So do the precisions: 4-bit here, 2–3-bit there. The cluster mostly buys 4-bit weights, not speed. Two Macs in tensor parallel decode at about 25 tok/s, not far above one M3 Ultra at 3-bit.

Could one Mac hold a 4-bit build? On paper, ggml-org’s Q4_K_S is 470.1 GB, which leaves about 40 GB of a 512 GB Mac for macOS and the KV cache. mlx-community’s oQ4e is 477.6 GB and Unsloth’s UD-Q4_K_XL is 505.5 GB. Hugging Face lists Q4_K_S among the quants it tried on Apple Silicon but published no speed or memory figure for it. Until someone does, treat single-Mac 4-bit as undemonstrated.

Why a 256 GB Mac Studio is not enough

Hugging Face’s launch post is direct: even the 3-bit build needs more than 256 GB of RAM. The smallest GGUF at 3 bits or more is Unsloth’s UD-IQ3_XXS at 340.4 GB. The index lists the model’s minimum as 342 GB, which puts it in the 512 GB tier. What does fit in 256 GB is below 3 bits:

A 256 GB Studio can load Muse Spark Open only at 1–2 bits, below the floor Unsloth recommends for agentic work, and no speed has been published at that size. If you own a 256 GB Mac, the index’s pick is MiMo-V2.6-Flash, the #1 model overall (Score 82; needs a 192 GB+ Mac Studio). It runs at community59 tok/s community-measured on a 256 GB M3 Ultra and estimated~79 tok/s estimated on an M5 Ultra. On a Mac with 32 GB or less, the pick is Qwen 3.6-35B-A3B (Score 81).

Day-zero runtimes and versions

Support landed upstream in both main Mac runtimes on launch day. Status checked on 18 October 2026:

# llama.cpp b11671 or newer: the launch post's command
llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K

# MLX: the re-uploaded oQ3e needs mlx-lm 0.32.2 or newer
pip install -U "mlx-lm>=0.32.2"
mlx_lm.generate --model mlx-community/Muse-Spark-Open-816B-A42B-oQ3e \
  --prompt "Explain this repository's build system" --max-tokens 5000

Plan for the download. The builds run 304–372 GB, which takes roughly 40–50 minutes on a sustained 1 Gb/s connection.

Muse Spark Open vs Muse Spark 1.2 and 1.3

Since August the index’s watch list had carried one Meta item: open weights for Muse Spark. Meta launched Muse Spark 1.2 with Muse Code in August and followed with 1.3 in September, both as API models, and had promised open weights without a date. Muse Spark Open is that release, under a new name rather than a version number.

Its lineage comes from Meta’s PyTorch blog. That post says Muse Spark Open shares its base model with Muse Spark 1.3 and adds safety post-training so it holds up as a standalone model without Meta’s system guardrails. The model card and Meta’s research blog mention the safety post-training but not the 1.3 base. So this is neither 1.2’s weights nor 1.3 itself, and the card’s benchmarks describe Muse Spark Open only. Muse Spark 1.2 and 1.3 stay API-only. Muse Spark Open is also served by API hosts, among them Baseten, Fireworks AI, Groq, OpenRouter and Together AI, according to Meta’s blog.

The watch item is closed. Meta’s small open model is unchanged: Muse Glimmer 30B (Apache 2.0, 18 GB minimum) runs at sourced27 tok/s published on an M5 Max.

Against the other 512 GB models

Muse Spark Open is the largest model in the index with a published single-Mac speed. It is not the only frontier-scale model a 512 GB Mac Studio can hold, though, and of the group below it is the hardest to fit. Its 42B active parameters are the most of the four. It is also the only one whose demonstrated single-Mac builds are below 4-bit.

Single-Mac figures from the LLMCheck index. Muse Spark Open’s M3 Ultra figure is at Q2_K and its M5 Ultra estimate at oQ3e. DeepSeek V4.1 Flash is at oQ4e, with the community run made on the base V4.1 checkpoint with MTP off. Ling-3.1-flash and Beam 1.0 are at Q4_K_M.

Read the speed columns with the quant in mind. Muse Spark Open’s M3 Ultra figure is at 2-bit and the others are at 4-bit. Per token it reads the most weights of the four, so at the same precision on the same Mac it would be the slowest; the 2–3-bit build is what keeps it level. Two of its rivals carry runtime caveats: DeepSeek V4.1 Flash runs only through oMLX, and Beam 1.0 has no published run on Apple Silicon. DeepSeek’s Mac path is covered in DeepSeek V4.1 Flash and MiMo-V2.6 on a Mac.

Should you buy a 512 GB Mac for it?

• You already own a 512 GB M3 Ultra. Run it. Start with the oQ3e build in oMLX or mlx-lm 0.32.2+ at low or medium effort. Hugging Face’s figure for the original upload is 21 tok/s (published, M3 Ultra).
• You are buying a Mac for this model. Wait for the 512 GB M5 Ultra, due in late October. The index puts it at estimated~31 tok/s estimated at oQ3e, in line with the 1.6× single-Mac decode gain over the M3 Ultra that Apple published for Qwen3.8-27B. A published M5 Ultra run will replace the estimate. The M5 Ultra Mac Studio explainer covers the rest of the machine.
• You want 4-bit. That is a two-Mac project: two 512 GB M5 Ultras over Thunderbolt 5, at 24.99 tok/s in Apple’s tensor-parallel test (published, multi-Mac).
• You want the best model a big Mac can run. The score says MiMo-V2.6-Flash, and Muse Spark Open ranks below it on both the 512 GB M3 Ultra and 512 GB M5 Ultra pages. Muse Spark Open’s case is breadth: audio and video input, the top MCP-Atlas score on its card, and a licence whose main obligation is a credit line.

Frequently Asked Questions