Meta's Muse Spark Open has about 816 billion parameters, and only about 42 billion are active for each token. The launch coverage calls it an open-weight model that "runs on Macs," but the memory it needs depends on how it's packaged. The tested figures range from about 300GB to 1TB, and the license adds a revenue threshold most open models don't have. For a founder, the question isn't whether the model runs on a Mac. It's whether the hardware, license, and volume make local hosting cheaper or safer than a hosted API.
Hardware: What It Actually Takes
The 1TB figure is one estimate, not a rule. Engadget reports that the model needs "a staggering 1TB of RAM," suggesting two 512GB M5 Ultra Mac Studios or four 256GB units. It estimates the four-unit option at roughly $38,000.
The launch documentation shows smaller setups running quantized versions. The Hugging Face launch post reports a 3-bit MLX version at about 344GB, running at roughly 21 tokens per second on an M3 Ultra with 512GB. A 2-bit llama.cpp build ran at 18 tokens per second using 304GB. The same page says the full-precision checkpoint is about 1.5TB.
The 512GB configuration is unconfirmed. Guru3d reports that M5 Ultra Mac Studios reach stores with up to 512GB, but Engadget says Apple hasn't announced pricing for the 512GB model. Apple also dropped the 512GB M3 Ultra option in March 2026. Check the current configurator before you budget.
Throughput is the missing number. None of the coverage I read gives tokens per second for the M5 Ultra. The only speeds I could verify are from the older M3 Ultra tests above.
Angle 1: Local Versus API Economics
Hardware cost has to be spread over real usage. A common method divides hardware cost by months of use, adds electricity and maintenance, and divides by monthly tokens. Using the $38,000 estimate with an assumed 24-month straight-line depreciation, that's about $1,583 per month before electricity and maintenance.
Break-even depends on the API you'd replace. Using GPT-4o-mini's listed prices of $0.15 input and $0.60 output per million tokens, a 4-to-1 input-to-output blend costs about $0.24 per million tokens. That puts break-even near 6.6 billion tokens per month. Against GPT-5-mini at $0.25 and $2.00, the threshold falls to about 2.6 billion.
The comparison assumes equal quality. The coverage I read gives no API price for Muse Spark Open, and Meta's benchmark claims are company-released. A cheaper API only matters if it does the same job.
For most founders, the answer is no. At these volumes, you'd need a large, steady workload before owning hardware makes sense. The calculation also ignores electricity, setup time, and the cost of keeping the system running.
Angle 2: Privacy and Data Control
Local hosting keeps prompts off third-party APIs. If your customers' data is sensitive, running the model on your own machine avoids sending it to a hosted provider. That matters most for healthcare, finance, and legal data.
Local doesn't remove compliance duties. Your own storage, access controls, and logging still apply. Running a model locally is one control among several.
Check whether you actually need it. If your data is public or your customers have accepted a provider's terms, the privacy gain may not justify the hardware.
Angle 3: The License
The license has a revenue threshold most open models lack. According to a summary of the Muse Model License, offering the model as an AI service with over $20 million in revenue over 12 months requires advance notice to Meta and a separate signed agreement. Internal use and building the model into your own products don't trigger it.
Attribution is required. Products using the model must display "Built with Muse Spark Open" in the interface, documentation, or settings. Apache 2.0 and MIT don't require this.
The acceptable use policy is stricter. It bans weapons, malware, critical infrastructure work, and fraud, and the summary says it also rules out military use and states the model isn't intended for anyone under 18.
This comes from secondhand summaries. The Hugging Face page doesn't include license terms, so read Meta's license text before you build on the model.
Angle 4: The Hardware Reality Check
The headline and the spec sheet don't match. "Runs on Macs" is true in the sense that Apple machines can run quantized versions. For a founder with a standard laptop, it isn't practical. The tested setups need more than 256GB of memory.
Plan for the cheapest setup that fits your use case. If you need a model for internal coding help, a smaller open model may cost less and run faster. Test the smaller option first.
Angle 5: Coding and Agent Workflows
Meta paired the model with a coding terminal. Meta says the model is paired with Muse Code, its AI coding terminal, and shows it working on a Mastodon client to fix bugs and build new features. Those are company demos, not independent results.
Benchmark claims are company-released. Meta says the model outperforms GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. The launch page doesn't publish scores or comparison details, so treat the claim as unverified.
Test before you switch. Run your own coding tasks on a current hosted model and on the open model, measure fix rates and time to completion, and compare them with the cost of each.
Recommendation Blocks
Consider local hosting if: - You have sustained high token volume, not occasional use - Your data can't leave your own systems - Your product's license terms are clear to you, and your revenue stays under the threshold Stick with a hosted API if: - Your volume is low or uneven - You can't justify hardware in the $30,000 to $40,000 range - You need a model you can start using this week
A Decision Framework
Step 1: Count your tokens. Measure monthly input and output volume for the tasks you'd move.
Step 2: Price the alternatives. Compare current hosted API prices for models that pass your quality test.
Step 3: Check the license. Read the full license text before committing, especially the revenue and attribution terms.
Step 4: Confirm the hardware. Check the current configurator and throughput before you assume a setup works.
Bottom Line
Muse Spark Open is a large open-weight model that runs on Apple hardware, but the practical cost is steep. Local hosting makes sense for high, steady volume and sensitive data. For most founders, a hosted API is cheaper and faster to start with.
Before you act on the headline, check three things: the memory configuration you'd buy, the license terms, and your own token volume. The rest is marketing until you test it.
