Meta Launches Muse Spark - Replaces Llama with Proprietary Superintelligence AI

Meta has just announced Muse Spark, their new flagship AI model that marks a dramatic shift from the open-source Llama family to a proprietary system. This is the first model from Meta Superintelligence Labs since the 2025 reorganization.

End of the Llama Era

Muse Spark represents Meta's exit from open-source AI. While Llama models have been available to everyone, Muse Spark is only accessible through the Meta AI app and a private API preview. This shift comes after Llama 4's disappointing launch and admitted benchmark manipulation issues.

Technical Breakthroughs

Visual Chain of Thought: Muse Spark is built from the ground up as a natively multimodal model capable of annotating dynamic environments and providing visual reasoning.

Contemplating Mode: A new mode that orchestrates multiple sub-agents for parallel reasoning, designed to compete with Google's Gemini Deep Think and OpenAI's GPT-5.4 Pro.

Thought Compression: The model achieves its reasoning capabilities using over an order of magnitude less compute than Llama 4 Maverick.

Benchmark Results

• Intelligence Index Score: 52 (vs Llama 4 Maverick's 18)
• CharXiv Reasoning: 86.4% (beats Claude Opus 4.6's 65.3%)
• HealthBench Hard: 42.8% (massive lead on health tasks)
• MMMU Pro: 80.4% (second-best vision model on market)

Practical Applications

• Shopping Mode: AI-powered recommendations based on Instagram and Threads content
• Health Reasoning: Analyzes food photos for nutrition and health scoring
• Interactive UI: Generates web-based minigames and tutorials

Future of Open Source

Meta's Alex Wang promises larger models are in development with plans for open-sourcing in the future, but many in the developer community remain skeptical of this reversal.

With Muse Spark, Meta has reintroduced itself as a direct challenger in the race for "Personal Superintelligence," though agentic workflows still represent a challenge.