Meta released Muse Spark on April 8, 2026, its first AI model under the newly formed Superintelligence Labs division — and its first significant departure from the Llama lineage that defined the company's open-source AI strategy for the past several years. This is not an incremental update. Internally codenamed "Avocado," Muse Spark is a ground-up rebuild: new architecture, new infrastructure, new data pipelines.
The timing and the personnel make the stakes clear. Alexandr Wang, who sold Scale AI to Meta in a $14.3 billion deal and now leads Superintelligence Labs, is the face of this release. Wang built Scale AI into the dominant AI data infrastructure company in the industry. His fingerprints on Muse Spark — particularly in data quality and evaluation methodology — are evident in the benchmark results, which are among the more interesting numbers from a model launch this year.
Muse Spark scores 52 on the Artificial Analysis Intelligence Index, placing it in the top five models benchmarked. It sits behind Gemini 3.1 Pro Preview and GPT-5.4, but ahead of Claude Sonnet 4.6, Grok 4.20, and several others. The benchmark that stands out most is HealthBench Hard, where Muse Spark scores 42.8 — above GPT-5.4 at 40.1 and well above Gemini 3.1 Pro at 20.6. In multimodal understanding, it scores 80.5% on MMMU-Pro, second only to Gemini 3.1 Pro's 82.4%.
What makes Muse Spark genuinely different from prior Meta AI releases is that it's a reasoning model — meaning it works through problems step by step, adjusts strategy when an approach isn't working, and can handle multi-agent orchestration natively. It takes in and outputs both text and images. For a company that spent the last two years playing catch-up to OpenAI and Anthropic on the product side, this is a meaningful shift in capability posture.
There's also an efficiency angle worth noting. Muse Spark completed the full Intelligence Index evaluation using 58 million output tokens — matching Gemini 3.1 Pro and dramatically below Claude Opus 4.6 (157 million) and GPT-5.4 (120 million). That matters for cost and latency at scale, which matters for the business.
The proprietary nature of Muse Spark — not open-source, not in the Llama family — signals a strategic pivot. Meta built its AI credibility on openness. Muse Spark suggests the company believes it now has something worth keeping proprietary. Whether that confidence is validated by real-world performance will become clearer as developers get access and start pushing the model against actual production workloads. But the early numbers give that confidence a foundation it hasn't always had.
Comments