Muse Spark Open Review: What the License, Hardware and Benchmarks Really Mean
Meta put an 816B open-weight model on Hugging Face. The license allows free use for most developers, but it carries a credit rule, a revenue threshold and a very large hardware bill.
Jump to 9 sections
Muse Spark Open is Meta's open-weight model, released October 15, 2026, with 816 billion parameters. The weights are free for many developers under a custom license, but you must credit Meta, services above $20 million in revenue need an agreement, and the hardware bill is steep. Published scores trail top Chinese open models.
This review is for developers and teams weighing a self-hosted frontier-class model. Model Drop tracks new models, platforms and tools, and this review reads the license, the hardware needs and the benchmark claims. We did not run our own test suite on it, so every score below is a published figure, and we say who published it.
What Did Meta Release?
Meta released the open-weight version of Muse Spark, the model it introduced in April. ITdaily's coverage of the open-weight release reported that the model comes from Meta Superintelligence Labs, has a one-million-token context window and focuses on software development and AI agents. Meta calls it the most powerful American open-weight model to date.
According to Trending Topics' report on the release, the model is natively multimodal across text, images and video. Weights are on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry. API access runs through Together AI, Fireworks, Groq and OpenRouter.
The context matters. Muse Spark, released in spring, was Meta's first model without open weights, a break from its Llama approach. This release partly reverses that, while the newer Muse Wonder models stay closed.
What Does the Muse Model License Allow?
The license is custom, and the details matter more than the word open. Meta describes it as having limited commercial terms that allow many developers to use the model for free. Free use, copying, distribution, fine-tuning, derivative models and distillation are all allowed, and Meta claims no rights to model outputs.
The limits are specific, as Trending Topics lists them. Offering the model as an AI service with more than $20 million in revenue over 12 months requires advance notice to Meta and a separate agreement. Internal use, research and embedding the model in your own products are exempt.
Two more rules apply. Any product using the model or a derivative must prominently display Built with Muse Spark Open. The license also bans uses such as weapons, malware, critical infrastructure attacks, fraud, disinformation and impersonation, and the model is not intended for users under 18.
| Question | Muse Model License | Typical Apache 2.0 model |
|---|---|---|
| Free commercial use | Yes, with limits | Yes |
| Credit required | Yes, visible notice | Attribution in notices only |
| Revenue threshold | $20 million for AI services | None |
| Fine-tuning and distillation | Allowed | Allowed |
The table above compares the license with the standard permissive one. Trending Topics notes that Alibaba's Qwen and Zhipu's GLM mostly use Apache 2.0 or MIT, which have no revenue thresholds or naming rules. Whether this counts as open source depends on your definition, and many developers will say it does not.
What Hardware Does It Need?
A lot. Trending Topics, citing Unite.AI, says the model needs at least about 700 GB of graphics memory, depending on format. Meta says it can run on two linked Mac Studios with M5 Ultra chips, which Trending Topics notes cost roughly as much as a compact car when configured with 512 GB each.
That is a specific kind of accessible. A hobbyist cannot run it, but a funded team or a small lab can. Mixture-of-experts models only activate 42 billion parameters per request, which helps speed, but all 816 billion must still sit in memory.
Most teams will use a host instead. The providers listed above will serve it by API, and prices were not given at launch. Our guide to inference providers compared explains why price should not be the first question when you choose one.
If you want to try local serving anyway, see our look at local AI inference servers compared for the tools that handle large models.
How Do the Benchmarks Look?
The numbers are Meta's, not an independent lab's. In Trending Topics' summary of Meta's table, Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, against 46 for MiMo-V2.6-Pro, the best Chinese model listed. On Terminal-Bench 2.1 it scores 76.4 against 85.0 for Kimi K3.
On GPQA Diamond, a science test, it scores 89.0 against 93.5 for Qwen3.8 and Kimi K3. On SWE-Atlas-QnA, a software test, it scores 58 against 66 for Kimi K3. It leads on MCP-Atlas, a tool-use test, at 89.2.
Against US and European peers it looks stronger. Meta's table shows it clearly ahead of Nvidia's Nemotron 3 Ultra and Thinking Machines' Inkling, often by more than 20 points. It is roughly tied with Mistral Large 4, which scores 38 on Artificial Analysis.
The independent ranking has not caught up. Trending Topics says Muse Spark Open is not yet on the Artificial Analysis open-weight list, and that Mistral Large 4 is absent because it is classified as proprietary. A full safety evaluation report is expected soon, according to ITdaily.
How Does It Fit With Meta's Other Models?
Meta now runs a split strategy. The closed Muse Wonder family sits at the top of its lineup and is sold through the Meta Model API and the Muse Code coding product. The open Muse Spark Open sits a tier below, and Meta lets the community host it.
That split mirrors what other labs do. A closed flagship earns revenue and sets the quality bar, while an open model builds goodwill, attracts developers and shapes standards. ITdaily notes that Meta has not said whether organizations can download and run Muse Wonder themselves.
For buyers, the practical message is that open does not mean top of the line. If you need the best score Meta has to offer, you pay for the closed models. If you need control and a lower bill, the open one is the choice, and you accept the gap.
The safety angle also matters. ITdaily reports that the model received additional safety training and that Meta says it meets the publication thresholds of its Superintelligence Scaling Framework. Because open weights cannot be recalled, a published safety report matters more here than for an API model.
Who Should Use Muse Spark Open?
Teams that need a US-origin open-weight model have a real option. Some companies cannot use Chinese models because of policy or customer rules. For them, the top of the open field has been thin, and Muse Spark Open fills that gap.
It also suits teams that want data control and agent work. Its tool-use score and one-million-token context fit long workflows. Self-hosting keeps data inside your own systems, which matters in regulated work.
It is a weaker choice for teams that want the best raw score per dollar. The Chinese models score higher on Meta's own table, and many carry plain Apache 2.0 or MIT terms. A startup planning to resell AI services should read the revenue threshold closely.
How Should You Evaluate It Before Committing?
Test it on your own tasks, as you would with any launch. Our explainer on what a point release actually changes is a reminder that headline numbers hide task-level differences. Run it on a set of real prompts next to your current model.
Use these checks before you commit.
- Read the full license text with your lawyer, not just the summary.
- Price the hardware or the hosted API for your expected volume.
- Run 20 to 50 real tasks against your current model.
- Check the credit notice against your product design.
- Wait for the independent ranking and the safety report.
The checks above cost little and prevent the most common surprises. They also give you numbers that matter for your product, not for a leaderboard.
Our Verdict
Muse Spark Open is a meaningful release for open-weight users, mostly because of who made it. It is the strongest US open-weight option Meta says exists, with a very long context and good tool-use numbers. The license is workable for most, and the hardware is a real barrier.
It is not the clear leader. On Meta's own comparison, four Chinese models score higher on the main index. We give it a conditional recommendation: try it if you need US origin, data control or agent features, and skip it if you only want the top score.
At Model Drop, we will update this review when Artificial Analysis lists the model and Meta publishes its safety report.
Conclusion
Muse Spark Open gives teams a large, multimodal, open-weight model from a US lab, with a custom license, a credit rule and a $20 million threshold for AI services. The hardware needs are heavy, and the published scores put it behind top Chinese models.
Your next step is to read the license, price your hosting options and test it on your own tasks. Revisit the decision once independent rankings and the safety report appear.
