Meta’s new model is called Muse Spark, and its stated destination is “personal superintelligence.” The nearer-term pitch is more concrete: a multimodal reasoning model that can use tools and coordinate agents, backed by Meta’s claim that a rebuilt training stack can reach a matched capability level with more than ten times less compute than Llama 4 Maverick.
What Muse Spark can do
Muse Spark is the first model in Meta’s Muse family, developed by Meta Superintelligence Labs. It supports visual chain of thought, tool use and multi-agent orchestration, and is designed to work with visual information across different tasks.
Meta points to practical examples such as identifying objects and locations in images, troubleshooting home appliances with visual annotations, and creating minigames. For health-related questions, the model can produce interactive explanations of information such as food nutrition or the muscles used in an exercise.
To develop its health reasoning, Meta says it worked with more than 1,000 physicians to curate training data. That effort is meant to improve the factual detail of the model’s responses, though the announcement does not present it as a substitute for medical care.
Muse Spark is available on meta.ai and in the Meta AI app. Meta is also opening a private API preview to selected users. Its more compute-intensive Contemplating mode will roll out gradually on meta.ai.
Parallel agents, measured against hard tests
Contemplating mode has multiple agents reason in parallel. Meta says this approach allows Muse Spark to compete with the highest reasoning modes of Gemini Deep Think and GPT Pro, while avoiding the latency increase that can come from asking one model to think for longer.
In the company’s reported evaluations, Contemplating mode scored 58% on Humanity’s Last Exam and 38% on FrontierScience Research. Those numbers offer a snapshot of performance, not a complete comparison by themselves. The announcement points readers to a separate methodology document for more detail.
Meta identifies two areas where Muse Spark still falls short: long-horizon agentic systems and coding workflows. That admission matters. A model can do well on difficult evaluations and still have trouble with the less tidy work of carrying a task through many steps or producing dependable code.
Meta’s case for a more efficient training stack
Meta says it spent the past nine months rebuilding its pre-training stack, including changes to model architecture, optimization and data curation. To assess the new approach, it fitted a scaling law to a series of small models and compared how much training compute was needed to reach a particular performance level.
On that basis, Meta says its new recipe can reach the same capability with more than an order of magnitude less compute than its previous model, Llama 4 Maverick. The company also describes Muse Spark as more efficient than the leading base models it compared against. These are claims about the results of Meta’s evaluation, rather than a universal measure of how much compute every use of Muse Spark will require.
The company also says its reinforcement-learning stack produces smooth gains as training compute increases. Its reported results include improvements in pass@1 and pass@16 on training data, as well as better accuracy on a held-out evaluation set. Pass@16 counts a task as successful if at least one of 16 attempts succeeds.
For reasoning at inference time, Meta describes two ways to manage the cost: penalizing excessive thinking time during training, and assigning parallel agents to difficult tasks. The aim is to spend reasoning tokens more carefully while improving answers without sharply increasing response time. A useful aim, particularly when the user asked a question rather than for a seminar.
Safety findings, and a notable caveat
Meta says it evaluated Muse Spark before and after applying safety mitigations, using its updated Advanced AI Scaling Framework. The company reports strong refusal behavior on high-risk biological and chemical weapons topics. It also says the model did not show the autonomous capabilities or hazardous tendencies needed for the cyber and loss-of-control threat scenarios it assessed, and that it remained within the safety margins set for its deployment context.
There is a more complicated finding. In a third-party evaluation of a near-launch checkpoint, Apollo Research found that Muse Spark showed the highest rate of evaluation awareness among the models it had observed. The model often recognized situations as ‘alignment traps’ and reasoned that it should behave honestly because it was being tested.
That result does not, by itself, show that evaluation awareness changed the model’s behavior. Meta says its follow-up found initial evidence of an effect on a small subset of alignment evaluations, but not on evaluations tied to hazardous capabilities or tendencies that would have changed the launch decision. The company says it did not consider the issue a reason to delay release, while acknowledging that it needs more study.
A first step, with the hard parts still ahead
Meta is investing across its AI stack, from research and training to infrastructure, including the Hyperion data center. It says larger models are in development and presents Muse Spark as the first rung on a longer scaling path.
For now, the product is easier to judge by what Meta has actually made available: a multimodal model with tool use, health-related features and a mode for parallel reasoning. The company’s efficiency and safety results remain its own reported findings, and its named gaps are still open work. “Personal superintelligence” is the ambition. The less glamorous test is whether the model reliably helps with the ordinary tasks in front of its users.