Meta’s Muse Spark Wraps Up Its Open-Source AI Venture
On April 8, Meta launched Muse Spark, its first entirely closed AI model, signaling a shift away from its open-source Llama initiative.
Summary
- Muse Spark, which debuted on April 8, is a product of Meta Superintelligence Labs, developed by Alexandr Wang’s team after a significant $14.3bn partnership with Scale AI.
- This model is fully proprietary with no public weights, diverging from the Llama strategy, which reached 1.2 billion downloads by early 2026.
- On the day of its launch, Meta’s stock rose by 9%, with plans to integrate the model into WhatsApp, Instagram, Facebook, and Messenger in the following weeks.
Meta revealed Muse Spark on April 8 as its first completely closed AI model, clearly moving away from its former open-source Llama strategy. This launch is the initial output from Meta Superintelligence Labs, a division established around Alexandr Wang following Meta’s hefty $14.3bn investment in Scale AI.
Wang remarked, “Nine months ago, we overhauled our entire AI framework. We built new infrastructure, innovative architecture, and restructured data pipelines. This is just the beginning; larger models are in development, with plans to open-source future versions.”
Unlike Llama, the weights for Muse Spark are not publicly accessible. Currently, API access is restricted to invited partners. Meta has expressed interest in potentially open-sourcing later versions, framing the current closure as temporary.
According to analyst Arun Chandrasekaran from Gartner, this transition marks a “significant shift,” indicating Meta’s clear intent to move away from the Llama branding entirely.
Functionality of Muse Spark
This model is inherently multimodal, capable of handling text, images, and voice inputs. Its notable feature is an operational mode called “Contemplating,” which allows multiple reasoning agents to run simultaneously before generating a response, effectively competing with Gemini Deep Think and GPT Pro.
Meta has collaborated with over 1,000 healthcare professionals to compile health-focused training data, promoting the model as a personal health reasoning tool while also functioning as a general assistant.
On the Artificial Analysis Intelligence Index, Muse Spark scores below GPT-5.4 and Gemini 3.1 Pro, achieving a score of 52 compared to their 57. Meta has not yet disclosed specifics regarding the model’s parameters or architecture. As reported by crypto.news, it demonstrated better performance than Gemini 3.1 Pro on various health metrics outlined in Meta’s evaluation criteria.
Reasons Behind Meta’s Timing
According to crypto.news, Meta has indicated a gradual approach towards its upcoming AI generation, opting to keep core elements proprietary while reviewing safety considerations.
The move to a wholly closed initial offering is viewed as a response to competitive pressures from OpenAI and Anthropic, both of which have proprietary models yielding billions in API revenue—an approach that Meta’s open-source strategy was unable to effectively exploit.
For 2026, Meta’s capital expenditures are anticipated to be between $115bn and $135bn, nearly double that of 2025. The stock saw a significant 9% increase on the launch day, marking the largest single-day response to a product announcement from Meta in over two years. The developer community that once thrived on Llama must now await a future open-source release, with no current timeline established.
