Meta Shifts from Llama to Proprietary Muse Spark

One billion downloads. That was Meta’s Llama milestone just a year ago. Today, the team at Meta had one billion downloads of their Llama project. The team behind Llama has since changed their focus and moved on to other projects.

They have redefined how they see AI by pivoting from an open source system to a proprietary one. In this article, we will provide a breakdown of what Muse Spark is, what it means to developers who built on Llama, and your choices going forward.

From open source champion to closed shop

Back in 2023, Meta released Llama 2 and positioned it as a win for the open-source community. CEO Mark Zuckerberg even published a personal essay in 2024 titled “Open Source AI is the Path Forward.” Developers took him at his word. They downloaded Llama over a billion times, built products on it, and trusted it as a long-term foundation.

Recently, Meta announced Muse Spark – its proprietary AI Model created by its newly formed Meta Superintelligence Labs. Meta Superintelligence Labs was created after Mark Zuckerberg hired Alexandr Wang, the creator of Scale AI, and is focusing on a new path of cutting-edge AI (frontier) development. Muse Spark was created from the ground up with a completely new architecture, infrastructure, and training pipeline. There are no similarities to Llama, right down to the architecture under the hood. 

What exactly is Muse Spark

Muse Spark, Meta’s answer to GPT-4 and Claude, is a closed, cloud-only model. It does not have download weights and cannot be used self-hosted or fine-tuned by outside developers. It is currently in a private API preview, which means most people cannot use it.

Meta has not officially declared Llama dead, but it has indicated that Llama’s existing models will remain available as open-source, albeit carefully worded. There is no comment from Meta as to future investments, development of new Llama models, or competition with other models such as Muse Spark. The consensus in the AI community is that Llama is moving to maintenance mode and that all of the significant resources of Meta will be used to develop Muse Spark.

Sources have indicated that Zuckerberg was frustrated watching Llama fall behind on benchmarks such as ChatGPT and Claude. Muse Spark was born out of that frustration.

What this means for Llama developers

This is where things get uncomfortable. There is no migration path from Llama to Muse Spark. The two systems are fundamentally different. Llama lets you download model weights, run them locally, and fine-tune them on your own data. Muse Spark does none of that. It lives in the cloud, behind an API, and Meta controls everything.

AI researcher Andrew Ng captured the mood well, noting that the proprietary release has worried many developers who built real projects on open-weights Llama models. Switching costs are real. Moving to a new model means rewriting APIs, adapting training workflows, and rebuilding tooling from the ground up.

Three paths forward for Llama users

If you are currently running Llama in production, here is what you can actually do. You can keep using your existing Llama models. They still work. Major cloud providers still host them. The honest downside is that these models will fall further behind frontier competitors with each passing month.

If you have not planned on using an open-weighted model and have an API as your primary requirement, then you can choose from any of the following companies and APIs: Claude, GPT-4, or Gemini, which have all been developed for you as fully functional options. For example, it was recently reported that part of Meta’s engineering team had already transitioned to using Claude Sonnet over Muse Spark before it was ever completed. 

There is also a fourth route worth knowing about. The llama.cpp ecosystem has produced several strong forks that work with Llama model files and extend well beyond Meta’s official support. Projects like ik_llama.cpp offer better CPU and hybrid GPU performance, while OpenLLaMA provides Apache-licensed reproductions of the original Llama models. These community tools will likely outlast Meta’s official interest in the architecture.

The bigger picture

Meta’s move is understandable from a business perspective. Proprietary models are easier to monetise, easier to control, and more attractive to enterprise customers who want reliability guarantees. But it does leave a gap in the open-source AI landscape that was previously filled by Llama’s scale and Meta’s backing.

For developers, the lesson is a familiar one. Building on any single company’s open-source commitment carries risk. What gets released freely today can be quietly deprioritised tomorrow.

If you are deciding where to build next, now is a good time to review your dependencies and think about which open-weights communities have sustainable development behind them. The Llama era is not fully over, but its best days are likely behind it.