Meta Ditches Llama, Launches Closed-Source 'Muse Spark' Flagship

Summary: Meta pivoted away from open-source Llama with Muse Spark, a proprietary flagship model offered only to select partners via private API — the first product of its newly formed Superintelligence Labs.

Key Points

• Closed API, no public weights: Unlike every Llama release, Muse Spark is fully proprietary — even more restrictive than paid API models from OpenAI or Anthropic.
• $14.3B rebuild from scratch: Meta hired Alexandr Wang (Scale AI founder) as Chief AI Officer and spent nine months dismantling and rebuilding its entire AI stack before shipping Muse Spark.
• Competitive benchmark position: Scores 52 on the Artificial Analysis Intelligence Index — behind GPT-5.4 and Gemini 3.1 Pro (both 57) and Claude Opus 4.6 (53), placing it in the global top five.
• Multimodal and agentic: Outperforms Llama 4 across multimodal perception, reasoning, healthcare, and agentic tasks.

Why It Matters

Meta was the loudest advocate for open-source AI; this reversal signals that frontier model training costs have crossed a threshold where open weights no longer make financial sense even for a company with Meta's resources. The shift also removes a major contributor from the open-source ecosystem at a critical moment.

Read More

• VentureBeat: Meta launches Muse Spark — VentureBeat
• The New Stack: Meta abandons Llama strategy — The New Stack