Muse Spark Open local LLM VRAM requirements

Muse Spark Open (Meta, open-weighted 2026-10-15) is an 816B-total / 42B-active multimodal MoE. Official BF16 siblings sum to 1,644,333,393,864 bytes (1,531.40 GiB, Meta says about 1.65 TB). Official NVFP4 siblings sum to 533,011,023,472 bytes (496.41 GiB, Meta says about 537 GB). The public model card says a GPU with at least 700 GB of VRAM for NVFP4 and 2 TB for BF16, and that the model is optimized to run on two Mac Studios with M5 Ultra using MLX. Our formula estimates 1,533.40 GiB–1,533.97 GiB (BF16) and 498.41 GiB–498.97 GiB (NVFP4) at 8,192 tokens, one stream. These are estimates, not measurements.

This result is an estimate, not a measured trace.Not verifiedFile-backed weights

Muse Spark Open GPU memory: official files versus Meta's floors

Meta open-weighted Muse Spark Open on 2026-10-15: 816B total / 42B active, 1M-token context, natively multimodal. The public model card says a GPU with at least 700 GB VRAM for NVFP4 and 2 TB for BF16, and that the model is optimized for two Mac Studios with M5 Ultra using MLX. Hugging Face sibling files (gated repo; index JSON was not readable without a license grant) sum to 1,644,333,393,864 bytes BF16 (1,531.40 GiB; Meta says about 1.65 TB) and 533,011,023,472 bytes NVFP4 (496.41 GiB; Meta says about 537 GB).

How to download and run it: see How to run Muse Spark Open locally for the weight files, gating and Meta's hardware note.

Single-device fit at 8,192 tokens, 1 stream, SGLang. Two GPUs or two Macs are never added into one VRAM pool.
Device	Usable	BF16 (1,531.40 GiB)	NVFP4 (496.41 GiB)
GeForce RTX 4090 24GB (desktop) 24 GB GDDR6X	22.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB
GeForce RTX 5090 32GB (desktop) 32 GB GDDR7	30.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB
NVIDIA L40S 48GB 48GB GDDR6 with ECC	46.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB
NVIDIA A100 80GB SXM 80GB HBM2e	78.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB
NVIDIA H100 80GB SXM 80GB	78.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB
NVIDIA DGX Spark (GB10, 128 GB unified) 128 GB LPDDR5x coherent unified system memory	112.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB
Apple M4 Max 128GB unified (MacBook Pro) 128 GB unified memory	116.0 GiB	Does not fit 1,533.40 GiB–1,533.97 GiB	Does not fit 498.41 GiB–498.97 GiB

Meta's two-Mac-Studio path uses Apple M5 Ultra unified memory (up to 512 GB per Studio, per the 2026-10-15 research post). This catalog has M2 Max 64 GB and M4 Max 128 GB only. Serving recipes on 8 × B300 (BF16) and 4 × B300 (NVFP4) are from Meta's Get the model page. No B300 SKU is in the catalog.

Memory by context and checkpoint

Checkpoint	2,048 tokens	8,192 tokens	32,768 tokens
Official BF16 safetensors	1,531.95 GiB–1,532.51 GiB	1,533.40 GiB–1,533.97 GiB	1,539.22 GiB–1,539.78 GiB
Official NVFP4 safetensors	496.95 GiB–497.51 GiB	498.41 GiB–498.97 GiB	504.22 GiB–504.78 GiB

Verified facts

ModelMuse Spark Open, meta-models/Muse-Spark-Open-816B-A42B
NVFP4 repometa-models/Muse-Spark-Open-816B-A42B-NVFP4
LicenseMuse Model License (gated Hugging Face form). Commercial Model-as-a-Service above $20M trailing revenue needs a separate Meta agreement.
ParametersCard: 816B total / 42B active. HF BF16 safetensors.total: 822,036,529,664
Language backbone62 layers, hidden 8192, GQA 128 Q / 16 KV, head dim 64, 256 experts top-8, vocab 202,048, 1M context
Perception Encoder1.8B, 50 layers, 1536 hidden, 16 heads (model card). Encoder bytes not separable (config.json gated).
ModalitiesInput: text, image, video, audio. Output: text.
Official GPU floorsAt least 700 GB VRAM (NVFP4) and 2 TB (BF16), plus two M5 Ultra Mac Studios via MLX
Unverified herePeak memory on B300 or Mac Studio, Unsloth GGUF sizes, any community 2-bit "fits 256 GB RAM" claim

How this page estimates an 816B MoE

Weights are sibling file-size sums because the safetensors index was gated. Meta's rounded TB/GB figures are shown next to those sums.
Experts stay resident. 42B active is compute, not a 42B memory footprint.
KV uses the published GQA shape (16 KV heads, head dim 64, 62 layers). Vision/audio encoder caches are not modelled.
Two Mac Studios are a Meta MLX topology. This calculator never adds two unified-memory machines into one pool unless a verified joint recipe exists.

Muse Spark Open FAQ

How much VRAM does Muse Spark Open need?

Meta's public model card: at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. Official Get the model docs list about 1.65 TB (BF16) and about 537 GB (NVFP4) of weights, excluding KV cache and runtime. Our file-backed sums are 1,531.40 GiB and 496.41 GiB. Formula demand at 8,192 tokens is 1,533.40 GiB–1,533.97 GiB (BF16) and 498.41 GiB–498.97 GiB (NVFP4).

Does Muse Spark Open need two Mac Studios?

Meta's 2026-10-15 research post says Muse Spark Open is the first model in its class to run on two Apple M5 Ultra Mac Studios with MLX, and that each Studio's unified memory is up to 512 GB. That is a publisher topology, not a catalog combo. This catalog has M2 Max 64 GB and M4 Max 128 GB SKUs only. Two discrete GPUs are never added into one VRAM pool.

Can it run on one RTX 4090, RTX 5090, or DGX Spark?

Not as a single unsplit replica of the official BF16 or NVFP4 checkpoints. BF16 files are 1,531.40 GiB; NVFP4 files are 496.41 GiB. Those exceed every single-device SKU in this catalog, including 80 GB H100/A100 and 128 GB unified Spark / M4 Max.

Is 816B or 822B the parameter count?

The model card says 816B total / 42B active. Hugging Face safetensors metadata on 2026-10-16 lists 822,036,529,664 parameters for the BF16 repo. This catalog stores the safetensors total for weight math and the card's 42B as active compute. All 256 experts stay resident unless a recipe offloads them.

Are the weights gated?

Yes. Hugging Face reports gated: manual. config.json and the safetensors index returned HTTP 401 without a license grant on 2026-10-16. Architecture fields in this catalog come from the public model-card HTML. File sizes come from the public HF API sibling list.

Publisher and provenance

PublisherMeta
Hugging Facemeta-models/Muse-Spark-Open-816B-A42B
Revision0fb02478b02565b2ce7fc57e6051c91172774ac9
Total parameters822,036,529,664
Active parameters42,000,000,000
Layers62
Query heads128
KV heads16 (gqa)
Head dimension64
Recorded context capability1,048,576
Completenesspartial · partial

Active parameters describe tokens' routed compute, not the full resident model weight footprint. Full-resident weight estimates use total parameters.

Memory layers in this estimate

Artifact bytes, raw KV, engine extras, and total demand are separate. This is not a runtime trace.

Weight artifact: 1,531.40 GiB (1,644,333,393,864 bytes)
Raw KV cache: 1.94 GiB (2,080,374,784 bytes)
Engine extras / workspace: 0.63 GiB–2.56 GiB
Total estimated demand: 1,533.97 GiB–1,535.90 GiB

Published artifacts

Quantization	Selected files	Bytes	Engine	Evidence
Official BF16 safetensors bf16 Calculator: Official BF16 safetensors	65 safetensors shards Total selected bytes: 1,644,333,393,864 (1,531.40 GiB / 1,644.33 GB) First shard: model-00001-of-00065.safetensors Last shard: model-00065-of-00065.safetensors View all 65 shard files `model-00001-of-00065.safetensors` `model-00002-of-00065.safetensors` `model-00003-of-00065.safetensors` `model-00004-of-00065.safetensors` `model-00005-of-00065.safetensors` `model-00006-of-00065.safetensors` `model-00007-of-00065.safetensors` `model-00008-of-00065.safetensors` `model-00009-of-00065.safetensors` `model-00010-of-00065.safetensors` `model-00011-of-00065.safetensors` `model-00012-of-00065.safetensors` `model-00013-of-00065.safetensors` `model-00014-of-00065.safetensors` `model-00015-of-00065.safetensors` `model-00016-of-00065.safetensors` `model-00017-of-00065.safetensors` `model-00018-of-00065.safetensors` `model-00019-of-00065.safetensors` `model-00020-of-00065.safetensors` `model-00021-of-00065.safetensors` `model-00022-of-00065.safetensors` `model-00023-of-00065.safetensors` `model-00024-of-00065.safetensors` `model-00025-of-00065.safetensors` `model-00026-of-00065.safetensors` `model-00027-of-00065.safetensors` `model-00028-of-00065.safetensors` `model-00029-of-00065.safetensors` `model-00030-of-00065.safetensors` `model-00031-of-00065.safetensors` `model-00032-of-00065.safetensors` `model-00033-of-00065.safetensors` `model-00034-of-00065.safetensors` `model-00035-of-00065.safetensors` `model-00036-of-00065.safetensors` `model-00037-of-00065.safetensors` `model-00038-of-00065.safetensors` `model-00039-of-00065.safetensors` `model-00040-of-00065.safetensors` `model-00041-of-00065.safetensors` `model-00042-of-00065.safetensors` `model-00043-of-00065.safetensors` `model-00044-of-00065.safetensors` `model-00045-of-00065.safetensors` `model-00046-of-00065.safetensors` `model-00047-of-00065.safetensors` `model-00048-of-00065.safetensors` `model-00049-of-00065.safetensors` `model-00050-of-00065.safetensors` `model-00051-of-00065.safetensors` `model-00052-of-00065.safetensors` `model-00053-of-00065.safetensors` `model-00054-of-00065.safetensors` `model-00055-of-00065.safetensors` `model-00056-of-00065.safetensors` `model-00057-of-00065.safetensors` `model-00058-of-00065.safetensors` `model-00059-of-00065.safetensors` `model-00060-of-00065.safetensors` `model-00061-of-00065.safetensors` `model-00062-of-00065.safetensors` `model-00063-of-00065.safetensors` `model-00064-of-00065.safetensors` `model-00065-of-00065.safetensors`	1,644,333,393,864 (1,531.40 GiB / 1,644.33 GB)	`sglang, vllm, unspecified`	`file-metadata`
Official NVFP4 safetensors nvfp4 Calculator: Official NVFP4 safetensors	127 safetensors shards Total selected bytes: 533,011,023,472 (496.41 GiB / 533.01 GB) First shard: model-00001-of-00127.safetensors Last shard: model-00127-of-00127.safetensors View all 127 shard files `model-00001-of-00127.safetensors` `model-00002-of-00127.safetensors` `model-00003-of-00127.safetensors` `model-00004-of-00127.safetensors` `model-00005-of-00127.safetensors` `model-00006-of-00127.safetensors` `model-00007-of-00127.safetensors` `model-00008-of-00127.safetensors` `model-00009-of-00127.safetensors` `model-00010-of-00127.safetensors` `model-00011-of-00127.safetensors` `model-00012-of-00127.safetensors` `model-00013-of-00127.safetensors` `model-00014-of-00127.safetensors` `model-00015-of-00127.safetensors` `model-00016-of-00127.safetensors` `model-00017-of-00127.safetensors` `model-00018-of-00127.safetensors` `model-00019-of-00127.safetensors` `model-00020-of-00127.safetensors` `model-00021-of-00127.safetensors` `model-00022-of-00127.safetensors` `model-00023-of-00127.safetensors` `model-00024-of-00127.safetensors` `model-00025-of-00127.safetensors` `model-00026-of-00127.safetensors` `model-00027-of-00127.safetensors` `model-00028-of-00127.safetensors` `model-00029-of-00127.safetensors` `model-00030-of-00127.safetensors` `model-00031-of-00127.safetensors` `model-00032-of-00127.safetensors` `model-00033-of-00127.safetensors` `model-00034-of-00127.safetensors` `model-00035-of-00127.safetensors` `model-00036-of-00127.safetensors` `model-00037-of-00127.safetensors` `model-00038-of-00127.safetensors` `model-00039-of-00127.safetensors` `model-00040-of-00127.safetensors` `model-00041-of-00127.safetensors` `model-00042-of-00127.safetensors` `model-00043-of-00127.safetensors` `model-00044-of-00127.safetensors` `model-00045-of-00127.safetensors` `model-00046-of-00127.safetensors` `model-00047-of-00127.safetensors` `model-00048-of-00127.safetensors` `model-00049-of-00127.safetensors` `model-00050-of-00127.safetensors` `model-00051-of-00127.safetensors` `model-00052-of-00127.safetensors` `model-00053-of-00127.safetensors` `model-00054-of-00127.safetensors` `model-00055-of-00127.safetensors` `model-00056-of-00127.safetensors` `model-00057-of-00127.safetensors` `model-00058-of-00127.safetensors` `model-00059-of-00127.safetensors` `model-00060-of-00127.safetensors` `model-00061-of-00127.safetensors` `model-00062-of-00127.safetensors` `model-00063-of-00127.safetensors` `model-00064-of-00127.safetensors` `model-00065-of-00127.safetensors` `model-00066-of-00127.safetensors` `model-00067-of-00127.safetensors` `model-00068-of-00127.safetensors` `model-00069-of-00127.safetensors` `model-00070-of-00127.safetensors` `model-00071-of-00127.safetensors` `model-00072-of-00127.safetensors` `model-00073-of-00127.safetensors` `model-00074-of-00127.safetensors` `model-00075-of-00127.safetensors` `model-00076-of-00127.safetensors` `model-00077-of-00127.safetensors` `model-00078-of-00127.safetensors` `model-00079-of-00127.safetensors` `model-00080-of-00127.safetensors` `model-00081-of-00127.safetensors` `model-00082-of-00127.safetensors` `model-00083-of-00127.safetensors` `model-00084-of-00127.safetensors` `model-00085-of-00127.safetensors` `model-00086-of-00127.safetensors` `model-00087-of-00127.safetensors` `model-00088-of-00127.safetensors` `model-00089-of-00127.safetensors` `model-00090-of-00127.safetensors` `model-00091-of-00127.safetensors` `model-00092-of-00127.safetensors` `model-00093-of-00127.safetensors` `model-00094-of-00127.safetensors` `model-00095-of-00127.safetensors` `model-00096-of-00127.safetensors` `model-00097-of-00127.safetensors` `model-00098-of-00127.safetensors` `model-00099-of-00127.safetensors` `model-00100-of-00127.safetensors` `model-00101-of-00127.safetensors` `model-00102-of-00127.safetensors` `model-00103-of-00127.safetensors` `model-00104-of-00127.safetensors` `model-00105-of-00127.safetensors` `model-00106-of-00127.safetensors` `model-00107-of-00127.safetensors` `model-00108-of-00127.safetensors` `model-00109-of-00127.safetensors` `model-00110-of-00127.safetensors` `model-00111-of-00127.safetensors` `model-00112-of-00127.safetensors` `model-00113-of-00127.safetensors` `model-00114-of-00127.safetensors` `model-00115-of-00127.safetensors` `model-00116-of-00127.safetensors` `model-00117-of-00127.safetensors` `model-00118-of-00127.safetensors` `model-00119-of-00127.safetensors` `model-00120-of-00127.safetensors` `model-00121-of-00127.safetensors` `model-00122-of-00127.safetensors` `model-00123-of-00127.safetensors` `model-00124-of-00127.safetensors` `model-00125-of-00127.safetensors` `model-00126-of-00127.safetensors` `model-00127-of-00127.safetensors`	533,011,023,472 (496.41 GiB / 533.01 GB)	`sglang, vllm, unspecified`	`file-metadata`
INT4 theoretical	No published file	411,018,264,832 (theoretical)	—	theoretical

Memory by quantization and context

Concurrent sequences = 1. KV defaults to FP16. The range includes documented runtime extras for the listed engine. These figures use the same core as the interactive calculator.

Quantization	2,048 tokens	8,192 tokens	32,768 tokens
Official BF16 safetensors	1,532.51 GiB–1,534.40 GiB	1,533.97 GiB–1,535.90 GiB	1,539.78 GiB–1,541.89 GiB
Official NVFP4 safetensors	497.51 GiB–499.40 GiB	498.97 GiB–500.90 GiB	504.78 GiB–506.89 GiB
INT4 theoretical Theoretical rows are not downloadable artifacts.	383.40 GiB–494.18 GiB	384.85 GiB–495.99 GiB	390.67 GiB–503.15 GiB

Raw KV contribution

Standard GQA/MHA raw KV is 2 × layers × kvHeads × headDim × tokens × sequences × bytes per element. This page uses KV heads (16), not query heads (128). Using query heads would overstate this cache by 8×.

8,192 tokens, 1 sequence, FP16: 2,080,374,784 bytes (1.9375 GiB)
32,768 tokens: 8,321,499,136 bytes (7.7500 GiB)

Open the calculator with this model

Sources

Muse Spark Open model card — Gated (manual). Card: 816B total / 42B active, 1M context, MoE transformer with Perception Encoder, 62 layers, GQA 128 Q / 16 KV, head dim 64, hidden 8192, 256 experts top-8, vocab 202,048. Open weights require at least 700 GB VRAM for NVFP4 and 2 TB for BF16. Optimized for two Mac Studios with M5 Ultra using MLX.
Hugging Face API (BF16) — sha 0fb02478b02565b2ce7fc57e6051c91172774ac9; gated manual; safetensors.total=822036529664; 65 safetensors siblings sum 1644333393864 bytes. config.json HTTP 401.
Hugging Face API (NVFP4) — sha 4239e4db19376904f2e82c167e78f70abb5455fe; 127 safetensors siblings sum 533011023472 bytes.
Meta Get the model — BF16 about 1.65 TB, TP8/EP8 on 8 × B300. NVFP4 about 537 GB, TP4/EP4 on 4 × B300.
Meta research blog — 2026-10-15. Two Apple M5 Ultra Mac Studios; 512 GB unified memory per Studio; 1.2 TB/s bandwidth quoted by Meta.
PyTorch blog — Open weights 2026-10-15 under the Muse Model License. 816B total / 42B active, 1M context.

Assumptions that affect this result

Requested context is within the recorded native window of 1048576 tokens.
Full-resident weight estimates use total parameters, including inactive MoE experts.
This MoE lists 822036529664 total parameters and 42000000000 active parameters. Active-parameter compute is not used as the weight-residency figure.
Weight bytes come from the selected weight files only, not from every file in the repository.
Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
Weight bytes come from 65 published safetensors shards totaling 1644333393864 bytes.
This checkpoint has a vision component. Text-only inference does not include vision encoder or projector bytes.
Context budget is the number of cached tokens per sequence (prompt plus reserved generation).

Calculator

Loading catalog snapshot…

Choose a model first, then check a device.

Model

Show historical models

Quantization

Device type

Hardware

Workload

Context length primarily affects KV cache memory.

Advanced settings

Runtime

Engine and KV cache type for this estimate.

EngineKV cache typeRecurrent / conv state dtype Include theoretical estimates (expert) Allow a substitute if the filter has no match

Workload

Overhead scenario and the prefill workspace.

Overhead scenario Include heavier prefill workspace

Topology

How the model is split across devices. Memory is not added together.

DeploymentIntra-replica splitDevice count

Replica count is an advanced deployment option. Quantity is nodes or GPUs, not replicas.

Replica groupsDevices per replica

Budget

System reserve and safety margin. An empty number keeps the previous value.

System / display reserve

GiB (2^30 bytes)

Budget already excludes system reserveExtra safety margin

Custom bits per weightDisplay unit

Custom model

Import a config or type the architecture. Nothing is scanned until you import.

Import local config.json

Parsed in the browser. The file is not uploaded.

Public Hugging Face repoRevision

Only huggingface.co, config.json, no token. Private repos are rejected.

Total parametersActive parametersLayersQuery headsKV headsHead dimension

Search or pick a catalog checkpoint to see memory demand.

Search or pick a catalog checkpoint to see memory demand.