Meta has effectively halted development of its open-source AI model Llama and shifted its strategy to the proprietary Muse Spark model, The News Stack reported on April 30 local time.

Meta in April unveiled Muse Spark, an AI model developed by Meta Superintelligence Labs, which it set up last year. Muse Spark was built from scratch on completely different infrastructure, architecture and data pipelines from Llama.

Meta said after unveiling Muse Spark that it will continue to provide existing Llama models as open source. But that is limited to maintaining the existing models, and it did not mention future development. The News Stack said the AI industry expects Meta to offer only small updates and maintenance for Llama.

Muse Spark differs fundamentally from Llama in how it is deployed, leaving no path at Meta to migrate to Muse Spark. Llama uses an open-weights approach that allows self-hosting and fine-tuning, while Muse Spark is cloud-only and does not allow weight downloads or self-hosting.

Llama had about 1.2 billion downloads as of a year ago. Thousands of companies and many developers built products based on Llama. AI expert Andrew Ng (Andrew Ng) said in his newsletter, "Meta abandoning its leading role in open-weight models is a big loss to the developer community."

Llama developers have three alternatives. They can keep using existing Llama models, but the technology gap with rival models is bound to widen. They can also switch to competing open-source models such as Mistral, DeepSeek and Alibaba Qwen. Another option is to move to proprietary APIs from major AI companies such as OpenAI, Google and Anthropic. Meta's internal developers are also said to have moved Llama to Claude Sonnet even before the launch of Muse Spark.