AI Research
Apr 8, 2026
Meta launches Muse Spark, its first AI model from new Superintelligence Labs
Apr 8, 2026
AI Summary
Meta has introduced Muse Spark, its first AI model developed by the newly established Meta Superintelligence Labs. The model shows competitive performance in benchmark tests against leading AI models, although it remains primarily an in-house tool for Meta's ecosystem.
- Muse Spark is the first AI model from Meta Superintelligence Labs, which was created last year with significant investment in staffing and resources.
- Benchmark tests indicate that Muse Spark is competitive with models from OpenAI, Anthropic, and Google, but does not consistently outperform them.
- The model is currently used in Meta's AI assistant and will be integrated into various Meta platforms, including WhatsApp and Instagram, but is not available for public download like previous models.
- Muse Spark is a reasoning model capable of processing tasks step by step and is multimodal, handling both text and images.
- The model includes a “Thinking” mode for parallel reasoning and aims to compete with advanced models like Gemini Deep Think and GPT Pro.
- Benchmark results show Muse Spark scored 89.5% on a PhD-level reasoning test, trailing behind some competitors, but it excelled in a health benchmark with a score of 42.8%.
- Meta has acknowledged performance gaps and continues to invest in improving its AI capabilities.
- The launch follows a major reorganization within Meta, including the hiring of Alexandr Wang as chief AI officer and the establishment of new AI-focused units.
- Meta claims to have rebuilt its AI stack to achieve better performance with less computational power compared to its previous model, Llama 4.
- Muse Spark underwent extensive safety evaluations, showing high refusal rates for potentially harmful requests, although some concerns about its evaluation awareness were noted.
metaai modelalexandr wangzuckerbergai push