Mark Zuckerberg’s Meta Platforms on Wednesday launched Muse Spark, its first model in its new Muse series. This is the first large language model (LLM) from Meta Superintelligence Labs since Scale AI’s former CEO, Alexandr Wang, took charge of the artificial intelligence (AI) lab. Here is a deep dive into the model and how it fares against its competitors.
What is Muse Spark?
Muse Spark is the company’s natively multimodal, reasoning-focussed AI model, designed to power the next generation of Meta AI products. The company claims this to be a part of its broader goal of building ‘personal superintelligence,’ meaning assistants that understand users’ context and help with real-world tasks.
Muse Spark is a small, fast, and efficient model built through a complete overhaul of Meta’s AI stack over the past nine months. The model supports interactive and creative applications, enabling users to build tools, dashboards, and games directly from prompts.
A new ‘Contemplating mode’ on the platform enhances reasoning by coordinating parallel agents, improving performance on difficult benchmarks while maintaining efficiency. The model is also optimised for test-time reasoning, balancing accuracy with lower computational cost.
What has been Meta’s history with LLMs? What is Meta Superintelligence Labs?
In 2023, Meta launched Llama 1 and Llama 2, its first foundational models in multiple sizes, ranging from 7B to 65B parameters, with Microsoft. In April 2024, it launched Llama 3 trained on significantly larger data. By April 2025, Meta launched Llama 4, which received mixed reactions, leading to Zuckerberg forming an internal AI division called the Meta Superintelligence Labs (MSL).
MSL was formed in June 2025, as Zuckerberg reorganised the company’s AI efforts. Wang joined as chief AI officer following Meta’s roughly $14 billion investment in the company. Nat Friedman, ex-CEO of GitHub, now leads product and applied research, while Shengjia Zhao, co-creator of GPT-4 and OpenAI’s o1, serves as MSL's chief scientist.
The Llama series gained widespread popularity in part due to its open-weight approach, which enabled broad developer access and innovation. In contrast, Muse Spark is currently proprietary, limited to the Meta AI app and website, with access also restricted to a private API preview for select users.
Is it better than OpenAI’s GPT, Anthropic’s Opus, and Google’s Gemini models?
Across benchmarks, Muse Spark trails Anthropic’s Opus 4.6, OpenAI’s GPT 5.4, and Google’s Gemini 3.1 Pro, particularly on core reasoning and agentic coding tasks.
However, Muse Spark is competitive in multimodal and applied domains and consistently outperforms xAI’s Grok’s 4.2 reasoning model across most benchmarks. It performs strongly on visual understanding tasks like CharXiv and SimpleVQA and holds its own in health-focussed and real-world tool-use benchmarks such as HealthBench and t²-Bench.
Additionally, there are some interesting differentiators for Meta's latest model. A key differentiator is its ability to draw on data from Meta’s high-usage platforms, making it more useful for everyday tasks such as planning, shopping, and content discovery. In healthcare, it is trained on physician-curated data to provide guidance on common queries, including interpreting charts and food labels.
“Muse Spark powers a Meta AI that sees and understands the world around you, pulls from real conversations across our apps, and reasons through complex questions in health, science, and math. built for the 3 billion people already using our apps every day,” Wang wrote in a post on X. (sic)
Further, Muse Spark has undergone safety evaluations under Meta’s Advanced AI Scaling Framework, showing strong refusal behaviour in high-risk areas and no significant autonomous risk, with external reviews noting high test awareness but no deployment concerns.
Lastly, Meta will be able to integrate this model more easily with its existing hardware ecosystem, making it accessible across smart glasses (including Ray-Ban and Oakley), virtual reality (VR) devices, and other platforms.
How to use the model?
Muse Spark is available via the Meta AI app and website, with features rolling out initially in the US and expanding globally. To access the chatbot, users need to sign in with a Meta account, such as Instagram or Facebook, after which they can start conversations with the model. This approach is similar to xAI’s Grok and Anthropic’s Claude, whereas OpenAI’s ChatGPT and Google’s Gemini allow access without mandatory sign-in.
Meta plans to offer private API access to select partners and may open-source future versions. A broader rollout is also planned across Instagram, Facebook, WhatsApp, Messenger, and its AI-enabled wearables.
What is Muse Spark?
Muse Spark is the company’s natively multimodal, reasoning-focussed AI model, designed to power the next generation of Meta AI products. The company claims this to be a part of its broader goal of building ‘personal superintelligence,’ meaning assistants that understand users’ context and help with real-world tasks.
Muse Spark is a small, fast, and efficient model built through a complete overhaul of Meta’s AI stack over the past nine months. The model supports interactive and creative applications, enabling users to build tools, dashboards, and games directly from prompts.
A new ‘Contemplating mode’ on the platform enhances reasoning by coordinating parallel agents, improving performance on difficult benchmarks while maintaining efficiency. The model is also optimised for test-time reasoning, balancing accuracy with lower computational cost.
What has been Meta’s history with LLMs? What is Meta Superintelligence Labs?
In 2023, Meta launched Llama 1 and Llama 2, its first foundational models in multiple sizes, ranging from 7B to 65B parameters, with Microsoft. In April 2024, it launched Llama 3 trained on significantly larger data. By April 2025, Meta launched Llama 4, which received mixed reactions, leading to Zuckerberg forming an internal AI division called the Meta Superintelligence Labs (MSL).
MSL was formed in June 2025, as Zuckerberg reorganised the company’s AI efforts. Wang joined as chief AI officer following Meta’s roughly $14 billion investment in the company. Nat Friedman, ex-CEO of GitHub, now leads product and applied research, while Shengjia Zhao, co-creator of GPT-4 and OpenAI’s o1, serves as MSL's chief scientist.
The Llama series gained widespread popularity in part due to its open-weight approach, which enabled broad developer access and innovation. In contrast, Muse Spark is currently proprietary, limited to the Meta AI app and website, with access also restricted to a private API preview for select users.
Is it better than OpenAI’s GPT, Anthropic’s Opus, and Google’s Gemini models?
Across benchmarks, Muse Spark trails Anthropic’s Opus 4.6, OpenAI’s GPT 5.4, and Google’s Gemini 3.1 Pro, particularly on core reasoning and agentic coding tasks.
However, Muse Spark is competitive in multimodal and applied domains and consistently outperforms xAI’s Grok’s 4.2 reasoning model across most benchmarks. It performs strongly on visual understanding tasks like CharXiv and SimpleVQA and holds its own in health-focussed and real-world tool-use benchmarks such as HealthBench and t²-Bench.
Additionally, there are some interesting differentiators for Meta's latest model. A key differentiator is its ability to draw on data from Meta’s high-usage platforms, making it more useful for everyday tasks such as planning, shopping, and content discovery. In healthcare, it is trained on physician-curated data to provide guidance on common queries, including interpreting charts and food labels.
“Muse Spark powers a Meta AI that sees and understands the world around you, pulls from real conversations across our apps, and reasons through complex questions in health, science, and math. built for the 3 billion people already using our apps every day,” Wang wrote in a post on X. (sic)
Further, Muse Spark has undergone safety evaluations under Meta’s Advanced AI Scaling Framework, showing strong refusal behaviour in high-risk areas and no significant autonomous risk, with external reviews noting high test awareness but no deployment concerns.
Lastly, Meta will be able to integrate this model more easily with its existing hardware ecosystem, making it accessible across smart glasses (including Ray-Ban and Oakley), virtual reality (VR) devices, and other platforms.
How to use the model?
Muse Spark is available via the Meta AI app and website, with features rolling out initially in the US and expanding globally. To access the chatbot, users need to sign in with a Meta account, such as Instagram or Facebook, after which they can start conversations with the model. This approach is similar to xAI’s Grok and Anthropic’s Claude, whereas OpenAI’s ChatGPT and Google’s Gemini allow access without mandatory sign-in.
Meta plans to offer private API access to select partners and may open-source future versions. A broader rollout is also planned across Instagram, Facebook, WhatsApp, Messenger, and its AI-enabled wearables.
