Goal has introduced spark musea proprietary AI model developed by its Superintelligence laboratory equipment. Unlike the Llama models, Muse Spark weights are not publicly accessible.
Currently, access is limited to the Meta AI portal at meta.ai, and API access is granted by invitation only. According to a blog post announcing the launch, Meta describes Muse Spark as “the first step in our growth ladder and the first product of a ground-up overhaul of our AI efforts.”
What is Meta’s Muse Spark model?
Meta describes the model as a native multimodal reasoning system equipped with tool usage, visual chain of thought, and multi-agent coordination features. The company has not revealed details about its underlying architecture or the number of parameters.
It has also introduced a “Contemplation mode” that executes multiple reasoning agents in parallel, with the goal of competing with Gemini deep thought and GPT Pro. This mode is not available at launch and is being gradually rolled out via meta.ai.
Muse Spark performance vs. other AI models
Meta claims that Muse Spark performs on par or exceeds benchmarks set by leading OpenAI, anthropicand Google. The company also claims that the model required significantly less computing to train compared to Call 4 while achieving similar levels of performance.
Meta has published its testing methodology along with benchmark results. These figures are representations of Meta himself. Earlier this year, Meta faced criticism for its benchmarking practices for the release of Llama 4.
Open Source Pivot and Meta Superintelligence Labs Roadmap
The closed statement signals a notable change from what Meta CEO Mark Zuckerberg described in a 2024 post titled “Open source AI is the way forward.” In that post, he argued that open source AI was the best way to achieve broad economic benefits and that releasing Llama did not threaten Meta’s business model.
Muse Spark follows Meta’s Llama API inference service, launched alongside Llama 4. This move marked a step towards monetizing access to these models. Zuckerberg also shared in a Threads post that Meta still intends to release open source models in the future alongside its proprietary options.
Muse Spark was developed by Meta Superintelligence Labs, a team formed after Meta stopped working on its largest variant of Llama 4, called Behemoth, which was planned to have 2 trillion parameters.
Meta has not revealed the size of Muse Spark or the other Muse variants already in development. Additional models in the Muse family have been confirmed to be in development, but no release dates have been announced.
