Meta has introduced a private AI model called Muse Spark, developed by the Superintelligence Labs team. Unlike the Llama models, the weights of Muse Spark are not publicly available.

Currently, access is limited to the AI portal of Meta at meta.ai, which provides API access by invitation only. According to a published blog post, Meta describes Muse Spark as the first step on our scaling ladder and the first product of a complete restructuring of our AI efforts.

What is Meta's Muse Spark Model?

Meta defines this model as a locally multimodal reasoning system equipped with features for tool usage, visual thought chaining, and multi-agent coordination. The company did not provide detailed information about the model's core architecture or the number of parameters.

Additionally, it introduced a Thinking Mode that aims to compete with Gemini Deep Think and GPT Pro, running multiple reasoning agents in parallel. This mode is not available at launch and is being gradually rolled out via meta.ai.

Muse Spark Performance Compared to Other AI Models

Meta states that Muse Spark performs at least as well as, or better than, the standards set by leading models from OpenAI, Anthropic, and Google. The company also claims that the model requires significantly less computational power for training compared to Llama 4 while achieving similar performance levels.

Meta has published its testing methodology and benchmark results. These figures represent Meta's own representation. Earlier this year, Meta faced criticism due to benchmark applications for the Llama 4 version.

Open Source Transition and Meta’s Superintelligence Labs Roadmap

The closed version marks a significant shift as described by Meta CEO Mark Zuckerberg in a 2024 post titled Open Source AI is the Way Forward. In that post, he argued that open source AI is the best way to achieve broad economic benefits and that the release of Llama did not threaten Meta's business model.

Muse Spark follows the launch of Meta's Llama API inference service, which was launched alongside Llama 4. This step is recorded as a move towards monetizing access to these models. Zuckerberg also shared in a Threads post that Meta still plans to release open source models alongside private options in the future.

Muse Spark was developed by Meta Superintelligence Labs, established after halting work on Behemoth, Meta's largest Llama 4 variant, which was planned to have 2 trillion parameters.

Meta has not disclosed the size of Muse Spark or other Muse variants currently under development. It has been confirmed that additional models in the Muse family are being developed, but no release date has been announced.