Muse Spark Open is an open-weight model. Meta’s research post on 15 October 2026 says it has 816B total parameters (42B active), a 1M-token context, and is natively multimodal. Meta says it runs on two Apple M5 Ultra Mac Studios with MLX. A single VRAM number for a discrete GPU is not published. This site does not calculate one. The estimator does not model this architecture. For the calculator’s per-GPU memory estimate, see the Muse Spark Open VRAM model page. This guide covers the other half: what you download, where it is gated, and what Meta says about running it.
Announced scale and the file metadata
| Item | Value | Source |
|---|---|---|
| Public release | 15 October 2026 | Meta AI Research |
| Parameters | 816B total, 42B active | Research post |
| Context | 1M tokens | Research post |
| Safetensors metadata, unsuffixed repo | 822,036,529,664 | HF API, revision 0fb02478b025 |
| Safetensors metadata, NVFP4 repo | 483,807,947,264 | HF API, revision 4239e4db1937 |
Repo createdAt |
15 September 2026 | HF API. This is repository creation, not the 15 Oct announcement |
| License, as stated in the post | Muse Model License, limited commercial terms; many developers can use it for free | Research post |
The 816B announcement and the 822,036,529,664 metadata total are not the same number. The NVFP4 metadata total is not an 816B parameter count. Checked 16 October 2026. The card text itself was gated.
Weight-file bytes
GB is 10⁹ bytes. GiB is 2³⁰ bytes. Each total is the sum of .safetensors files in that repo. Config, tokenizer, and index files are not included. There is one shard layout in each repo, so the shards are not a second copy of the same weights.
| Repo | Safetensors files | Bytes | GB | GiB |
|---|---|---|---|---|
meta-models/Muse-Spark-Open-816B-A42B |
65 | 1,644,333,393,864 | 1,644.333 | 1,531.405 |
meta-models/Muse-Spark-Open-816B-A42B-NVFP4 |
127 | 533,011,023,472 | 533.011 | 496.405 |
These are download sizes of the weight files. They are not runtime memory. KV cache, activations, and the framework are extra. See file size is not runtime VRAM.
The two-machine sentence
Meta writes that each M5 Ultra integrated GPU can use up to 512 GB, so two machines can hold an entire model, and that the chip’s memory bandwidth is 1.2 TB/s. Meta also calls the pair a terabyte of unified memory. That is Meta’s description of the machines. It does not name a revision or a quantization.
This site does not add two unified-memory computers into one pool, and it does not treat unified memory as discrete VRAM plus RAM. See two GPUs and unified memory. Two marketed 512 GiB pools are 1,024 GiB. The unsuffixed shard sum above is 1,531.405 GiB, which is larger than that pair. The NVFP4 shard sum is 496.405 GiB, which is under one 512 GiB label and still is not a measured runtime fit. No Apple M5 Ultra SKU is in this catalog, so there is no hardware page to link.
Meta names these engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Hosted inference partners named in the post are Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter. Those are availability notes, not memory measurements.
Not the Muse Spark API
The closed Muse Spark API is a different product. On Meta’s pricing page, standard mode for muse-spark-1.3, muse-spark-1.2, and muse-spark-1.1 is $0.15 cached input, $1.25 input, and $4.25 output per 1M tokens. That price is not a price for Muse Spark Open. Muse Wonder is the closed frontier API, and its weights are not this download.
What this page does not claim
No calculator preset and no will-it-fit page. Qwen3.8-27B is an example of a model the calculator does estimate. DGX Spark is a 128 GB unified-memory device in the catalog. It is linked so the unified-memory rule is clear. It is not a Muse Spark Open fit. Formula limits are on the methodology page.
FAQ
How much memory does Muse Spark Open need?
Meta says two Apple M5 Ultra Mac Studios can hold the model, and that each M5 Ultra integrated GPU can use up to 512 GB. Meta does not name the checkpoint in that sentence. The unsuffixed Hugging Face repo's 65 safetensors shards, revision 0fb02478b025, sum to 1,644,333,393,864 bytes (1,644.333 GB, 1,531.405 GiB). That is larger than 2 times 512 GiB. The NVFP4 repo's 127 safetensors shards sum to 533,011,023,472 bytes (533.011 GB, 496.405 GiB). This site did not measure a runtime fit, and it does not add two machines into one memory pool.
How many parameters does Muse Spark Open have?
Meta's 15 Oct 2026 research post says 816 billion total and 42 billion active, with a 1 million token context. The unsuffixed repo's safetensors metadata total is 822,036,529,664 parameters. The NVFP4 repo's metadata total is 483,807,947,264, which is a packed count, not 816 billion. Both numbers are reported. The repo creation timestamp 15 Sep 2026 is not the public release date.
Are the Muse Spark Open weights downloadable?
Meta says the weights are available on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. The Hugging Face repos are gated (manual). The model card text was not readable without access on the check date. License terms quoted here are only what the research post states: the Muse Model License with limited commercial terms, and many developers can use it for free. A revenue cap was not in that post.
Is Muse Spark Open the same product as the Muse Spark API?
No. Muse Spark Open is the open-weight release. Muse Spark 1.3, 1.2, and 1.1 are separate closed API model ids. Their standard token prices are on Meta's pricing page and are not a price for the open weights.