Meta’s Muse Spark Open Explained: The 816B Open-Weight Model That Runs on Two Mac Studios
Meta has released Muse Spark Open, an 816-billion-parameter open-weight model it calls the most capable in the United States. The twist: with Apple’s help, it can run locally on a pair of Mac Studios. Here’s what it is, what the license really allows, and who should care.
What’s in this guide
What Meta Announced
On October 15, 2026, Meta launched Muse Spark Open, an open-weight version of its Muse model family. TechCrunch reports that Meta positions it as the most capable open-weight model from a U.S. company, built for complex coding and agentic work.
Unusually, the launch came with Apple as a partner: the model is optimized to run on Apple’s MLX framework on Mac Studio machines with the M5 Ultra chip. Meta also paired it with Muse Code, its AI coding terminal.
Key Specs
| Spec | Muse Spark Open |
|---|---|
| Total parameters | 816 billion |
| Active parameters per request | ~42 billion (mixture-of-experts) |
| Context window | 1 million tokens |
| Inputs | Text, images and video (natively multimodal) |
| Focus | Coding and long-running agent tasks |
| Local hardware | Two linked Mac Studios (M5 Ultra), Apple MLX |
Specs as reported by TechCrunch and SQ Magazine. Memory needs depend on the format/quantization used.
Can You Really Run It at Home?
Technically yes — practically, it’s a serious workstation project. Running the full model requires roughly 700 GB to 1 TB of fast memory depending on format. That’s why the reference setup links two Mac Studios, each with up to 512 GB of unified memory. Reports describe the cost of that setup as comparable to a compact car.
For most people and teams, the easier route is the cloud. The weights are available on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and hosted API access is offered by providers such as Together AI, Fireworks, Groq and OpenRouter.
What the License Allows (and Doesn’t)
“Open-weight” doesn’t mean “no strings attached.” According to SQ Magazine’s review of the Muse Model License:
- Allowed: free use, fine-tuning, building derivative models, distillation and commercial products. Meta claims no rights over outputs.
- Revenue threshold: companies providing the model and earning above a set annual revenue threshold (reported as about $20 million) must notify Meta and sign a separate agreement. Internal use and research are exempt.
- Attribution: products must show “Built with Muse Spark Open” somewhere visible, such as the interface, docs or settings.
- Prohibited uses: weapons development, malware and fraud; it isn’t intended for users under 18.
If you plan to build a commercial product on it, read the full license yourself — the revenue clause is the part most likely to matter.
How Good Is It?
Meta says Muse Spark Open beats other popular open-weight models, but most early numbers come from Meta’s own charts. Independent comparisons suggest it is strong but not the top open model overall — leading Chinese open models such as Kimi K3 and Qwen3.8 still score higher on several coding and science benchmarks.
| Benchmark | Muse Spark Open | Best Chinese open model |
|---|---|---|
| Terminal-Bench 2.1 (coding) | 76.4 | 85.0 (Kimi K3) |
| GPQA Diamond (science) | 89.0 | 93.5 |
| SWE-Atlas-QnA | 58 | 66 (Kimi K3) |
Reported figures; independent verification was limited at launch.
Where it stands out is the combination: a U.S.-built model, permissive commercial use, a 1M-token context and an official path to local deployment.
Who Should Care
Privacy-sensitive teams
Run a frontier-class model fully on-premises without sending data to a third party.
Developers
Fine-tune or distill it into smaller, cheaper models for your own product.
Agent builders
Long context and agentic tuning suit multi-step coding and research agents.
Cost-conscious users
Hosted open-weight APIs are often cheaper than closed frontier models.
FAQ
Is Muse Spark Open free?
The weights are free to download and use, including commercially, but large providers above the license’s revenue threshold need a separate agreement with Meta.
What hardware do I need to run it locally?
Roughly 700 GB–1 TB of memory. Meta and Apple’s reference setup uses two linked Mac Studios with M5 Ultra chips.
Where can I try it without buying hardware?
Through hosted APIs such as Together AI, Fireworks, Groq and OpenRouter, or cloud platforms like Hugging Face, SageMaker, Google Cloud and Microsoft Foundry.
Is it better than Chinese open models?
Not on every benchmark. Early figures show models like Kimi K3 ahead in coding and science tests.
Related Reading on FutureLume
- Meta Muse Explained: The No. 1 AI Agent App — Features, Pricing, and Whether You Should Trust It
- GPT-6 Astra vs Claude Fable 5.1 vs Gemini 3.8 Flash vs Muse Spark 1.3: September 2026’s AI Model Launches Compared
Muse Spark Open is a big step for U.S. open-weight AI: powerful, commercially usable and genuinely runnable on local hardware — if you have the budget. Most teams should test it through a hosted API first and read the license carefully before building on it.
Sources: TechCrunch · SQ Magazine · AppleInsider
