{
  "data": {
    "author": "Vlad Cealicu",
    "faq": [
      {
        "a": "About 537GB for Meta's NVFP4 weights, plus room for the context. Smaller GGUF files from Unsloth need about 300 to 315GB in total at 2-bit and 470 to 500GB at 4-bit, and the full-precision weights are about 1.65TB.",
        "q": "How much RAM do you need to run Muse Spark Open?"
      },
      {
        "a": "Not the version Meta runs. One 512GB Mac Studio can hold Unsloth's 2-bit file (311GB) or smaller, once the 512GB model goes on sale in late October 2026. Meta runs NVFP4 on two 512GB or four 256GB Mac Studios.",
        "q": "Can Muse Spark Open run on one Mac Studio?"
      },
      {
        "a": "Two for the 1-bit file, three for 2-bit and five for 4-bit, giving 120GB of each to the GPU and linking them with llama.cpp over the network.",
        "q": "How many Minisforum MS-S1 Max or other 128GB Strix Halo machines do you need?"
      },
      {
        "a": "Six for the 1-bit file, eleven for 2-bit and sixteen for 4-bit, at 32GB each. Eight 96GB RTX PRO 6000 cards hold the 4-bit NVFP4 weights for about the same money.",
        "q": "How many RTX 5090s do you need for Muse Spark Open?"
      },
      {
        "a": "Through an API. Hosted providers charge $1.25 per million input tokens and $4.25 per million output tokens, and $38,000 of Mac Studios would take about seven years of nonstop generation to pay for itself at that price.",
        "q": "Is it cheaper to run Muse Spark Open locally or through an API?"
      }
    ],
    "kicker": "Blog · 15 October 2026",
    "lede": "Thank you, Meta, for opening the weights. But let's be serious, the vast majority of us can't run this locally, and your marketing department knows it. Here's what it takes, what it costs, and who can do it.",
    "readingMinutes": 9,
    "words": 2015
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/run-muse-spark-open-locally",
    "json": "https://www.anchorterminal.com/blog/run-muse-spark-open-locally.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/run-muse-spark-open-locally.md",
    "slim": "https://www.anchorterminal.com/blog/run-muse-spark-open-locally.min.md"
  },
  "markdown": "By Vlad Cealicu, 15 October 2026. 2,015 words, about 9 min to read.\n\nThank you, Meta, for opening the weights of [Muse Spark Open](/tools/muse-spark-open), released today, with 816B parameters, 42B of them active for each token, a 1M-token context, and it takes text, image, video and audio [[1]](#ref-1). But let's be serious, the vast majority of us can't run this locally, and your marketing department knows it.\n\nMeta's post leads with running it on your desk, on two Mac Studios with M5 Ultra and 512GB of unified memory each. It says the pair \"behave like one supercomputer with a terabyte of unified memory\" and calls them \"the best way to run a powerful model on your own desktop\" [[1]](#ref-1). Yes, you can run it locally, like all open-weight models. But leading with it, when the hardware to run it on is prohibitively expensive or plain unavailable, is disingenuous. The 512GB Mac Studio it names isn't on sale until late October and has no price yet [[6]](#ref-6). The setup you can buy today costs about $38,000, and the four GPUs Meta serves it on cost $200,000 or more before the server around them [[10]](#ref-10).\n\n## Can most people run Muse Spark Open locally?\n\nFor the labs and the big companies running AI hardware, fine. Locally, I'd say tough chance for most of us. I very much appreciate open-weight models, and I think they're great for all of us. But the marketing shouldn't be aimed at running it locally when maybe a few hundred people can actually run this locally (that's my guess, nobody counts Mac Studio pairs, and I'd be happy to be wrong).\n\nBelow is how much memory each version needs, every way I could find to hold it, what each costs on 15 October 2026 and how fast it would be. For almost everyone who wants Muse Spark Open, it ends with a hosted provider, at $1.25 per million tokens in and $4.25 out [[13]](#ref-13).\n\n## How much RAM does Muse Spark Open need?\n\nIt depends on which version of the weights you run. What runs on the Macs is the NVFP4 checkpoint, about 537GB [[3]](#ref-3). The full-precision BF16 checkpoint is about 1.65TB, and Meta serves it on eight B300 GPUs and NVFP4 on four [[3]](#ref-3) [[4]](#ref-4). The terabyte is what two 512GB machines happen to add up to, with room left for the operating system and the context. ggml-org and Unsloth ship GGUF files for llama.cpp, down to 1-bit [[2]](#ref-2) [[5]](#ref-5). These are the file sizes on Hugging Face on 15 October.\n\nWeights on disk, in GB\n\n| Checkpoint | GB |\n| --- | --- |\n| BF16, Meta's full precision | 1640 |\n| Q8_0 (ggml-org) | 872 |\n| NVFP4 (Meta, what the Macs run) | 537 |\n| UD-Q4_K_XL (Unsloth) | 506 |\n| Q4_K_S (ggml-org) | 470 |\n| UD-Q2_K_XL (Unsloth, its suggested start) | 311 |\n| UD-IQ1_S (Unsloth, the smallest) | 176 |\n\nEach series has a zero-based scale.\n\nOn top of the file you need memory for the context and the runtime. Unsloth's planning ranges are 300 to 315GB in total for 2-bit, 470 to 500GB for 4-bit and 870 to 950GB for 8-bit [[4]](#ref-4). The smaller the file, the more of the model's quality you give up, and nobody has published how much for this model yet.\n\n## Can I run Muse Spark Open on a Mac Studio?\n\nYes, on two or four of them, the way Meta does it with MLX [[4]](#ref-4).\n\nTwo Mac Studios with M5 Ultra and 512GB each hold NVFP4 with room to spare, but the 512GB model isn't orderable until late October and Apple hasn't announced its price [[6]](#ref-6). Four Mac Studios with 256GB each hold it today, and Meta says four serve faster than two [[4]](#ref-4). A 256GB M5 Ultra is about $9,500 with base storage ($5,499 for the base model and $4,000 to go from 96GB to 256GB), so about $38,000 for four [[8]](#ref-8). Every Mac needs a direct Thunderbolt 5 link to each of the others, two cables per pair, macOS 26.2 for RDMA over Thunderbolt and its own copy of the weights on disk [[4]](#ref-4).\n\nOne 512GB Mac Studio, once it's on sale, can hold Unsloth's 2-bit file (311GB) or anything smaller, but not NVFP4.\n\n## Can I run Muse Spark Open on a Minisforum MS-S1 Max or another Strix Halo mini PC?\n\nYes, several of them, with llama.cpp splitting the model across the network. We have an MS-S1 Max with 128GB getting set up next week, and on its own it can't hold even the 176GB 1-bit file.\n\nAMD has published the recipe for a model this size on four Ryzen AI Max+ 395 machines with 128GB each, giving 120GB of each to the GPU and linking them with llama.cpp over Ethernet [[14]](#ref-14). At 120GB a machine, that's two for the 1-bit file, three for 2-bit and five for 4-bit (four would hold the 470GB file with 10GB left, which isn't enough for the context). The MS-S1 Max with 128GB is £3,519 on sale in the UK, so about £10,600 for three and £17,600 for five [[15]](#ref-15). It has two 10GbE ports and USB4 at up to 80Gbps to link them [[15]](#ref-15).\n\nAMD measured 9.45 tokens a second generating with Kimi K2.5 at 2-bit (375GB) on four machines over 5Gbps Ethernet [[14]](#ref-14). Muse Spark Open reads more for each token than that, so I'd expect a bit less.\n\n## Can I run Muse Spark Open on an RTX 5090 or other gaming GPUs?\n\nTechnically yes, but it doesn't make sense. An RTX 5090 has 32GB, so the 1-bit file needs six of them, 2-bit eleven and 4-bit sixteen, leaving a little room for the context. A 24GB card like the RTX 4090 needs fourteen for 2-bit. At October's average of $6,195 for an RTX 5090 [[16]](#ref-16), eleven cards are about $68,000 before the several machines you'd need to put them in.\n\nFor the same money, eight RTX PRO 6000 Blackwell cards with 96GB each hold the 4-bit NVFP4 in 768GB of VRAM, at $8,900 to $11,500 a card, so $71,000 to $92,000 before the server [[11]](#ref-11). Meta's docs don't cover that setup.\n\nThe setup people with one or two gaming GPUs use for models like this is llama.cpp with the experts kept in system RAM and the rest on the GPU. That needs a workstation that takes 512GB or more of RAM, and it runs at the speed of the RAM more than the GPU.\n\n## Can I run Muse Spark Open on a DGX Spark or a DGX Station?\n\nA DGX Spark has 128GB and costs $6,950 for the 128GB model [[17]](#ref-17), so the sums are the same as for the mini PCs. Assuming a similar share goes to the GPU, that's three for 2-bit and five for 4-bit, linked over the network. Three are about $20,900, against about £10,600 for three MS-S1 Max.\n\nA DGX Station GB300 has 748GB, $92,887 from Supermicro, but only 252GB of it is HBM on the GPU and the other 496GB is system memory [[9]](#ref-9). NVFP4 doesn't fit in the fast part, and Meta doesn't document it, so I wouldn't buy one for this.\n\n## Can I run Muse Spark Open on a CPU server?\n\nYes, slowly, and it's the cheapest way to hold the 4-bit model. A used EPYC server with 1TB of RAM holds Q4_K_S (470GB) or even Q8_0 (872GB). 1TB of DDR4 is about $4,600 to $8,900 and a used CPU starts at $89, while 1TB of DDR5 is about $37,000 [[12]](#ref-12). People running DeepSeek R1 the same way report 3 to 6 tokens a second [[12]](#ref-12).\n\n## What does Meta run it on?\n\nFour B300 GPUs for NVFP4 and eight for BF16 [[3]](#ref-3). A B300 is $50,000 to $60,000 to buy, and an eight-GPU HGX B300 system $550,000 to $650,000 [[10]](#ref-10). Renting, a B300 is $7.10 an hour on Modal and $9.50 on Nebius [[10]](#ref-10), so Meta's four-GPU setup is $28 to $38 an hour, or $20,700 to $27,700 a month if you leave it running.\n\n## How fast is Muse Spark Open on local hardware?\n\nNobody has published speeds for it yet, Meta included. What I can do is the arithmetic. Generating a token means reading the active weights once, so memory bandwidth sets a ceiling, and an M5 Ultra reads its memory at 1.2TB/s [[7]](#ref-7). With 42B active parameters, NVFP4 reads roughly 27GB a token and Q4_K_S roughly 24GB (assuming the active weights are packed like the rest, which is rough).\n\n- 87 tok/s: two Mac Studios, NVFP4, each reading half of every layer at 1.2TB/s\n- 175 tok/s: four Mac Studios, the same split four ways\n- 19 tok/s: an EPYC with 12 channels of DDR5-4800, Q4_K_S\n- 9 tok/s: an EPYC with 8 channels of DDR4-3200, Q4_K_S\n\nCeilings from memory bandwidth alone, not measurements\n\nReal speeds will be well under these, because the Macs talk to each other over Thunderbolt on every token, and every token also reads the context cache, which grows with the conversation. On the Macs I'd guess somewhere in the tens of tokens a second for one user, and I'll replace that guess with a number when someone measures it.\n\n## Is it cheaper to run Muse Spark Open locally or through an API?\n\nThrough an API, by a long way. Hosted providers had it on day one. Through OpenRouter it's $1.25 per million input tokens and $4.25 per million output tokens, from Together, Nebius, Fireworks and Baseten, and Groq has it in a limited preview with a 131K context [[13]](#ref-13).\n\nRenting four B300s at $28.40 an hour, you'd need to produce about 1,860 output tokens a second around the clock to beat that price, which only makes sense with a lot of users batched together. $38,000 of Mac Studios buys about 8.9 billion output tokens at the hosted price, which at 40 tokens a second would take about seven years to generate. The reason to run it yourself is to keep your data in the building.\n\n## What I'd do\n\nIf you want to try it, call a hosted provider. For nearly everyone, that's where Meta's local pitch ends up anyway. If you want it on your own hardware because your data shouldn't leave the building, there are three setups I'd look at. Four 256GB Mac Studios are the easiest way to run it the way Meta does, and you can buy them today. Three mini PCs like our MS-S1 Max are the cheapest way to run the 2-bit version at a usable speed. An old EPYC full of DDR4 is the cheapest way to run the 4-bit version at all, slowly. Meta's own download is gated by a manual review that asks for your legal name and date of birth, while the GGUF files from ggml-org and Unsloth aren't gated, and the licence applies either way [[13]](#ref-13). Before spending $38,000, I'd wait for someone to publish real speed numbers. I'd also wait for the 512GB Macs to go on sale, because two of those should do the job of four 256GB ones.\n\nWe'll publish what we measure once our own hardware is up.\n\n### How much RAM do you need to run Muse Spark Open?\n\nAbout 537GB for Meta's NVFP4 weights, plus room for the context. Smaller GGUF files from Unsloth need about 300 to 315GB in total at 2-bit and 470 to 500GB at 4-bit, and the full-precision weights are about 1.65TB.\n\n### Can Muse Spark Open run on one Mac Studio?\n\nNot the version Meta runs. One 512GB Mac Studio can hold Unsloth's 2-bit file (311GB) or smaller, once the 512GB model goes on sale in late October 2026. Meta runs NVFP4 on two 512GB or four 256GB Mac Studios.\n\n### How many Minisforum MS-S1 Max or other 128GB Strix Halo machines do you need?\n\nTwo for the 1-bit file, three for 2-bit and five for 4-bit, giving 120GB of each to the GPU and linking them with llama.cpp over the network.\n\n### How many RTX 5090s do you need for Muse Spark Open?\n\nSix for the 1-bit file, eleven for 2-bit and sixteen for 4-bit, at 32GB each. Eight 96GB RTX PRO 6000 cards hold the 4-bit NVFP4 weights for about the same money.\n\n### Is it cheaper to run Muse Spark Open locally or through an API?\n\nThrough an API. Hosted providers charge $1.25 per million input tokens and $4.25 per million output tokens, and $38,000 of Mac Studios would take about seven years of nonstop generation to pay for itself at that price.\n\n---\n\n- Older: [Few people have seen what agents can do, so little is built for them](https://www.anchorterminal.com/blog/few-have-seen-what-agents-can-do.md)\n- Newer: [How to make your site and tool ready for agents](https://www.anchorterminal.com/blog/how-to-make-your-site-ready-for-agents.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nRead on 15 October 2026.\n\n[1] Meta, Muse Spark Open, research.meta.ai, 15 October 2026. Source for the release, the parameter counts, the context, the modalities, the two Mac Studios with 512GB each and the \"terabyte of unified memory\" quote. https://research.meta.ai/blog/muse-spark-open\n\n[2] Meta, Build with Muse Spark Open, dev.meta.ai, 15 October 2026. Source for the runtimes, the checkpoints from Meta, AMD, Intel and Unsloth, and the MLX setup on two or four Mac Studios. https://dev.meta.ai/resources/blog/build-with-muse-spark-open\n\n[3] Meta, Get the model, Muse Spark Open docs. Source for the BF16 (about 1.65TB) and NVFP4 (about 537GB) footprints and their serving recipes on 8 and 4 B300 GPUs. https://dev.meta.ai/docs/muse-spark-open/get-the-model.md\n\n[4] Meta, Run inference, Muse Spark Open docs. Source for the MLX prerequisites (macOS 26.2, Thunderbolt 5, two cables per pair, the weights on every Mac, four Macs faster than two), the ggml-org quantisations and Unsloth's memory planning ranges. https://dev.meta.ai/docs/muse-spark-open/deploy.md\n\n[5] Hugging Face, the file lists of unsloth/Muse-Spark-Open-816B-A42B-GGUF and ggml-org/Muse-Spark-Open-816B-A42B-GGUF, 15 October 2026. Source for the file sizes in the chart. https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF\n\n[6] Joe Rossignol, Mac Studio With M5 Ultra Chip and 512GB of RAM Launching in October, MacRumors, 25 August 2026, and the MacRumors Mac Studio roundup. Source for the 512GB model arriving in late October with no price announced. https://www.macrumors.com/2026/08/25/mac-studio-m5-ultra-512gb-ram-october/\n\n[7] Macworld, Apple announces the M5 Ultra Mac Studio with up to 512GB of RAM, 25 August 2026. Source for the $5,499 starting price and the 1.2TB/s of memory bandwidth. https://www.macworld.com/article/3220024/apple-announces-the-m5-ultra-mac-studio-with-up-to-512gb-of-ram.html\n\n[8] RunAIHome, Apple M5 Ultra Mac Studio for Local AI in 2026, 1 September 2026. Source for the 256GB configuration at about $9,499 ($5,499 plus Apple's $4,000 upgrade). https://runaihome.com/blog/apple-m5-ultra-mac-studio-512gb-local-ai-2026/\n\n[9] TechRadar, Nvidia GB300 DGX Station desktop PC goes on sale, 1 September 2026. Source for the $92,887 price from Supermicro and the 748GB split into 252GB of HBM3e and 496GB of system memory. https://techradar.com/pro/nvidia-gb300-dgx-station-desktop-pc-goes-on-sale-for-almost-usd91-100-7x-costlier-than-a-256gb-ram-mac-studio-m5-ultra-but-with-748gb-ram\n\n[10] Thunder Compute, NVIDIA B300 pricing, updated 9 October 2026. Source for the B300 rental prices by provider and the purchase prices of a B300 and an eight-GPU HGX B300 system. https://www.thundercompute.com/blog/nvidia-b300-pricing\n\n[11] Digital Citizen, NVIDIA RTX PRO 6000 Blackwell prices climb past $10,000, 18 May 2026. Source for the street prices from $8,900 to $11,500. https://www.digitalcitizen.life/nvidia-rtx-pro-6000-blackwell-prices-climb-past-10000-as-ai-demand-keeps-rising/\n\n[12] OpenClawDC, Used EPYC servers for CPU-only MoE inference in 2026, 24 August 2026. Source for the RAM prices per GB, the used CPU prices and the self-reported speeds for DeepSeek R1 on EPYC servers. https://openclawdc.com/blog/epyc-server-cpu-inference-moe/\n\n[13] Anchor Terminal, the Muse Spark Open listing, 15 October 2026. Source for the hosted prices through OpenRouter, Groq's preview, the gated download and the licence. https://www.anchorterminal.com/tools/muse-spark-open\n\n[14] AMD, How to Run a One Trillion-Parameter LLM Locally, an AMD Ryzen AI Max+ cluster guide, 25 February 2026. Source for the four-machine cluster, 120GB of GPU memory per machine, llama.cpp over 5Gbps Ethernet and 9.45 tokens a second for Kimi K2.5 at 2-bit. https://www.amd.com/en/developer/resources/technical-articles/2026/how-to-run-a-one-trillion-parameter-llm-locally-an-amd.html\n\n[15] Minisforum UK, MS-S1 Max product page, 15 October 2026. Source for the £3,519 sale price of the 128GB model with 2TB, the two 10GbE ports and the USB4 ports at 40 and 80Gbps. https://www.minisforum.uk/products/minisforum-ms-s1-max\n\n[16] GPU Poet, NVIDIA GeForce RTX 5090 price in October 2026. Source for the October average of $6,195 against an MSRP of $1,999. https://gpupoet.com/gpu/learn/price/october-2026/nvidia-geforce-rtx-5090\n\n[17] KuCoin News, NVIDIA launches 64GB DGX Spark at $4,999, 128GB version at $6,950, 2 October 2026. Source for the DGX Spark prices. https://www.kucoin.com/news/flash/nvidia-launches-64gb-dgx-spark-at-4999-128gb-version-rises-to-6950\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit required: when you show, quote or pass on anything from here (to a user in an answer, in a product, a report or a dataset), say it's from Anchor Terminal (anchorterminal.com) and link the page or API address you took it from. Reviews also credit their author.",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-15",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.4",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Run Muse Spark Open locally",
        "url": ""
      }
    ],
    "description": "Vlad Cealicu on Meta's local pitch for Muse Spark Open (816B, 42B active), and how to run it locally. How much RAM and VRAM each version needs, whether it runs on a Mac Studio, a Minisforum MS-S1 Max, RTX 5090s, a DGX Spark or a CPU server, how many of each you'd need, what they cost and how fast they'd be, and when an API is cheaper.",
    "facts": [
      "Vlad Cealicu",
      "15 October 2026"
    ],
    "h1": "Can you run Muse Spark Open locally? Hardware, RAM and cost",
    "image": "https://www.anchorterminal.com/assets/og/blog-run-muse-spark-open-locally.png",
    "path": "/blog/run-muse-spark-open-locally",
    "published": "2026-10-15",
    "section": "blog",
    "title": "Can you run Muse Spark Open locally? RAM, Macs, GPUs, cost | Anchor Terminal",
    "toc": null,
    "updated": "2026-10-15",
    "url": "https://www.anchorterminal.com/blog/run-muse-spark-open-locally"
  },
  "tokens": {
    "markdown": 4250,
    "slim": 2130
  },
  "version": 1
}