Meta launches Muse Spark to power chatbot, shopping push

Meta launched Muse Spark on April 8, a new AI model from Meta Superintelligence Labs that will power the Meta AI chatbot.

This marks a shift from the company’s earlier open-source approach to a closed model.

The release is an early test of CEO Mark Zuckerberg’s AI overhaul after Meta invested US$14 billion in Scale AI.

Muse Spark was built in nine months and trained partly on open models from Alibaba, as well as models from OpenAI and Google.

Meta is also considering paid API access and future subscriptions for Meta AI, while the model is already being used in an early shopping agent, the executive said.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Muse Spark’s performance leap and Meta’s move away from open source

Independent benchmarks rate Muse Spark far above Meta’s earlier models 1.
Artificial Analysis, an AI model benchmarking firm, gave Muse Spark 52 on its Intelligence Index. That places it among the five best models it has tested. Meta’s Llama 4 Maverick scored 18 at its April 2025 launch 2.
Meta rebuilt its pretraining stack, the system used to prepare a model before fine-tuning 2. The company says it matches Llama 4 Maverick capabilities with over an order of magnitude less compute, which equals more than 10 times less processing power 2. The release uses proprietary (non-open-weight) distribution and a private API (application programming interface) preview for select partners 2.

Meta is weaving specialized AI features into its products

Meta is leaning on health use cases and worked with more than 1,000 physicians to curate training data 1.
Muse Spark can create interactive displays that explain nutritional content or muscles used during exercise 1. Meta frames the model around reasoning plus real-world usefulness in science, math, and health 1.
A shopping-oriented mode is under development. It will draw from creator content across Instagram and Threads to deliver personalized product recommendations 1.