TL;DR
- New Model: Meta Superintelligence Labs has launched Muse Spark, its first AI model built under new chief AI officer Alexandr Wang.
- Capabilities: The multimodal model powers Meta AI across its apps with reasoning modes, health features, and visual coding tools.
- Competitive Position: Meta acknowledges Muse Spark is not state of the art but scores in the top 5 on independent benchmarks.
- Strategy Shift: The company plans to monetize the model through API access while also releasing open-source versions.
Meta Superintelligence Labs has shipped its first AI model, Muse Spark, nine months after Mark Zuckerberg spent $14.3 billion rebuilding the company’s AI program from the ground up.
The launch marks the first tangible output of Meta Superintelligence Labs, the division Zuckerberg created after investing $14.3 billion in Scale AI and recruiting its CEO Alexandr Wang as chief AI officer. The model, code-named Avocado, was built after the delayed and disappointing release of Llama 4 prompted a wholesale rethinking of Meta’s AI strategy.
Wang’s team rebuilt Meta’s AI stack from the ground up, recruiting researchers from OpenAI, Anthropic, and Google with compensation packages that Wired described as worth hundreds of millions.
What Muse Spark Offers
Muse Spark now powers the Meta AI assistant in the Meta AI app and Meta AI website in the United States. Meta describes the model as “purpose-built for Meta’s products,” drawing a parallel to Google’s Gemini integration across Google’s own suite. In coming weeks, it will roll out across WhatsApp, Instagram, Facebook, and Messenger, as well as Meta’s smart glasses, and expand to other countries.
The model is small and fast by design, yet capable of reasoning through complex questions in science, math, and health. It is natively multimodal, trained to handle images, audio, and video as well as text, though it currently produces text-only output. Users can choose between an Instant mode and a Thinking mode for more thoroughly reasoned results, similar to reasoning features from competitors. A more advanced Contemplating mode to tackle complex problems is planned for a future release, positioning the model for use cases that currently favor frontier systems.
One of the more distinctive technical features is Muse Spark’s ability to deploy multiple AI agents simultaneously on a single problem to improve speed without sacrificing reasoning quality. Meta also emphasizes health reasoning as a key differentiator: the company collaborated with over 1,000 physicians to curate training data, and the model can process health questions involving images and charts. In a demo, Meta showed the model estimating a calorie count for a meal, though such features remain hit-or-miss across the industry.
Beyond health, the model excels at visual coding, letting users generate custom websites and mini-games from a text prompt. A shopping mode draws from styling inspiration and brand storytelling across Meta’s apps, reflecting the company’s effort to tie the AI assistant directly into its commerce ecosystem.
All versions of Muse Spark are free to use, though Meta may impose rate limits. Users need a Meta account (Facebook or Instagram) to access it.
The product-first approach reflects a strategic shift for Meta. Rather than competing on raw model benchmarks, the company is betting that deep integration across its platform ecosystem, which according to Meta spans 3.3 billion users, gives Muse Spark a distribution advantage that standalone AI providers cannot match. The emphasis on health, shopping, and visual creation suggests Meta is prioritizing practical consumer use cases over the developer-focused tooling that defines competitors like OpenAI and Anthropic.
Competitive Standing
Meta is being unusually candid about where Muse Spark falls in the AI pecking order.
“Muse Spark doesn’t mark a new state of the art, but is competitive with the latest models from leading labs at certain tasks, including multimodal understanding and processing health information.”
A Meta executive (via Axios)
The company also acknowledges a gap between Muse Spark and existing models in coding tasks, one of the areas where developers are particularly likely to evaluate a new AI system. Meta’s own benchmark scores suggest the model outperforms competitors from OpenAI, Anthropic, Google, and xAI at certain tasks, but self-reported benchmarks have been a sore point for the company since Llama 4 faced allegations of benchmark manipulation. Independent benchmarking offers a more favorable assessment.
“Muse Spark scores 52 on the Artificial Analysis Intelligence Index, placing it within the top 5 models we have benchmarked.”
Artificial Analysis
The timing adds pressure. Anthropic detailed its latest Mythos AI model the same week, a model it says is so powerful its initial release is limited to a handful of tech companies. OpenAI’s model code-named Spud is nearing completion and believed to represent a notable leap forward. OpenAI’s ChatGPT Health and Anthropic’s Claude for Healthcare both debuted in January 2026, giving them a head start in the health AI space Meta is targeting.
The candid admission of limitations positions Meta to manage expectations while buying time for larger Muse models. For Zuckerberg, the gamble is that shipping an honest, product-integrated model now builds more credibility than waiting for a frontier-class system that may take years to arrive.
The path to this launch was turbulent. Meta found early success with its open Llama models but lost momentum with Llama 4, which faced reports of quality issues and benchmark manipulation. The company then retreated from open-source AI development in favor of proprietary models. Yann LeCun, Meta’s longtime chief AI scientist, quit in March after clashing with Wang, calling him “young and inexperienced,” adding to the sense of internal upheaval.
Roadmap and Strategy
Meta says the next Muse generation is already in development and plans to release a version of Muse Spark under an open-source license, signaling that the closed-source approach may be temporary. The company has also published an Advanced AI Scaling Framework document outlining its vision for safely scaling AI to superhuman levels, framing Muse Spark as a step toward Zuckerberg’s stated goal of personal superintelligence.
The company is exploring a new AI model revenue stream by offering third-party developers API access to Muse Spark’s technology, starting with a private preview for select partners. This would mark a departure from Meta’s tradition of giving away its AI models entirely for free, and signals the company sees commercial potential beyond advertising revenue.
Zuckerberg spelled out the broader vision on Threads: “We are building products that don’t just answer your questions but act as agents that do things for you.” Meta hopes Muse Spark will eventually power features that cite recommendations and content shared across Instagram, Facebook, and Threads, drawing on the company’s vast social graph as a source of personalized context that rivals cannot replicate.
The dual strategy of charging for API access while promising open-source versions mirrors an emerging pattern in the AI industry, where companies like Mistral and xAI offer both proprietary and open-weight models. For Meta, the approach could help offset the $14.3 billion investment while maintaining the developer goodwill that made Llama popular in the first place.
Muse Spark is the first model in the new Muse series, making it Meta’s second major AI model family after Llama. Meta itself calls it an “early data point” on the trajectory of the series, and the company describes it as “the first step on our scaling ladder.” Whether that ladder leads Meta back to the frontier of AI development will depend on how quickly the larger Muse models close the gap with rivals.