Muse Spark Open vs paying for an AI subscription

updated 2026-10-15 · prices checked on vendor pages

quick answer: Muse Spark Open is Meta's new open-weight model, released on october 15, 2026. the weights are free to download from Hugging Face, but running it at home takes two Mac Studios with M5 Ultra and 512 GB of memory each, or four with 256 GB each. for nearly everyone, a $20/mo subscription or paying per token through a hosted API works out far cheaper. running it yourself makes sense if you need the model on your own hardware, not to save money.

we haven't tested or scored Muse Spark Open yet. everything below comes from Meta, Hugging Face, OpenRouter and Apple coverage, linked at the bottom.

Muse Spark Open at a glance

what Meta actually released

Meta calls Muse Spark Open "the most capable US open-weight model to date". it's the open-weights version of Muse Spark, with extra safety training to make it harder to misuse when it runs without other guardrails. Meta says it's built for coding and agent work: long tasks, multi-file software changes and multi-step reasoning.

worth knowing: on the benchmark table in Meta's own model card, a few other open-weight models score higher on some tests. on the Artificial Analysis Intelligence Index, Meta lists Muse Spark Open at 39, against 44 for Kimi K3 and 45 for GLM 5.3. so "most capable US open-weight model" is a narrower claim than "best open model".

can you really run it on a Mac Studio?

yes, but you need a pair of very well-specced ones. Mark Zuckerberg posted on launch day that it "runs great on new Mac Studios thanks to work with @johnternus and Apple". Meta's hardware docs spell out what that means:

• two Mac Studios with M5 Ultra and 512 GB of unified memory each, or four with 256 GB each
• the NVFP4 checkpoint (about 537 GB of weights), run with Apple's MLX
• the Macs linked directly over Thunderbolt 5, running macOS 26.2 or later

not on a Mac? the model card says you'll need at least 700 GB of GPU memory for NVFP4, or 2 TB for BF16. Meta's recipes use four NVIDIA B300 GPUs for NVFP4 and eight for BF16.

the sums: own hardware vs a subscription

the software is free. the hardware very much isn't.

• the M5 Ultra Mac Studio starts at $5,499 with 96 GB of memory, which isn't enough on its own (Macworld).
• the 512 GB version is due in late october and Apple hasn't published its price yet. MacRumors expects it to start well above $10,000.
• four Mac Studios at the $5,499 starting price come to nearly $22,000, before the memory upgrade you'd need.

compare that with what a subscription costs (prices from our own listings):

at $20 a month, a subscription costs $240 a year. nearly $22,000 of Macs is over 90 years of that. even a heavy user is better off paying per token through a hosted API, unless you've got reasons beyond price.

when running it yourself does make sense

• your data can't leave your building. Meta's docs say that once the weights are downloaded everything runs offline, with no API key.
• you want to fine-tune it. the licence allows you to modify, fine-tune and distil the model, and Meta points to PyTorch's TorchTitan for tuning.
• you're building a product on it. the weights are free to use commercially, with conditions (below).
• you already own the hardware. then the sums look very different.

want to try it without buying anything?

use a hosted API. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. on OpenRouter it costs $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 (checked october 15, 2026). OpenRouter says its API is OpenAI-compatible, so most SDKs work by swapping the base URL.

for comparison, Meta's Muse Wonder is $4 / $20 per million tokens on the Meta Model API.

the licence, in plain english

the Muse Model License is free of charge and lets you use, modify, fine-tune, distil, share and sell copies of the model. the catches:

• you must show Built with Muse Spark Open in your product, its docs, website or settings.
• if you sell access to AI models as a service (via API or similar) and your company makes over $20 million a year, you need a separate agreement with Meta first. internal use and research aren't affected.
• there's a usage policy listing what you can't use it for, and it isn't meant for anyone under 18.

FAQ

is Muse Spark Open free?

the weights are free to download from Hugging Face under the Muse Model License. running it isn't free: you need serious hardware, or you pay per token on a hosted API such as OpenRouter ($1.25 / $4.25 per million input / output tokens).

how big is Muse Spark Open?

816B total parameters with 42B active per token, and a 1M-token context window. the weights are about 1.65 TB in BF16 or about 537 GB in NVFP4.

can Muse Spark Open run on a Mac?

yes, on two Mac Studios with M5 Ultra and 512 GB of unified memory each, or four with 256 GB each, using MLX over Thunderbolt 5 on macOS 26.2 or later.

is running Muse Spark Open cheaper than ChatGPT Plus or Claude Pro?

not for most people. nearly $22,000 of Mac Studios is over 90 years of a $20/mo subscription. run it yourself if you need it on your own hardware, not to save money.

tools in this guide

• OpenRouter
• Meta Muse Wonder
• Muse by Meta
• Claude Code
• OpenAI Codex
• Cline

sources

checked 2026-10-15

• Meta announcement: https://research.meta.ai/blog/muse-spark-open
• model card and licence (Hugging Face): https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B
• Meta developer blog: https://dev.meta.ai/resources/blog/build-with-muse-spark-open
• Meta docs, get the model: https://dev.meta.ai/docs/muse-spark-open/get-the-model
• Meta docs, hardware platforms: https://dev.meta.ai/docs/muse-spark-open/hardware-platforms
• Mark Zuckerberg on X: https://x.com/finkd/status/2110778025629347999
• OpenRouter pricing for Muse Spark Open: https://openrouter.ai/meta/muse-spark-open
• Macworld, M5 Ultra Mac Studio pricing: https://www.macworld.com/article/3220024/apple-announces-the-m5-ultra-mac-studio-with-up-to-512gb-of-ram.html
• MacRumors, 512 GB Mac Studio due late october: https://www.macrumors.com/2026/08/25/mac-studio-m5-ultra-512gb-ram-october/
• subscription prices: https://claude.com/pricing; https://developers.openai.com/codex/pricing; https://www.meta.com/help/subscriptions/1021145227643680/

more guides