Fordel StudiosMeta debuts Muse Spark AI model from Superintelligence Labs
What Happened
Meta released Muse Spark, the first model from its Superintelligence Labs division led by Alexandr Wang, on April 8, 2026. The model, developed codenamed Avocado, scores 86.4% on figure understanding benchmarks but 42.5% on abstract reasoning tasks. The release follows Meta's $14 billion investment in Wang's team and highlights that perception and reasoning capabilities remain difficult to develop in tandem.
Our Take
$14 billion and the abstract reasoning score is 42.5%. Let that sink in.
Look, 86.4% on figure understanding is legitimately impressive — that's the kind of score that makes multimodal pipelines actually useful for document parsing, UI analysis, chart reading. We could actually use this for something.
But here's the thing: abstract reasoning at 42.5% means it's going to fall apart the moment you ask it to do anything that requires more than pattern recognition. Coding agents, multi-step planning, anything where the model needs to actually *think* — forget it. It's a perception model pretending to be a reasoning model.
Alexandr Wang is sharp, his team is clearly not incompetent, and Meta's not short on compute. This tells you something uncomfortable about where the ceiling might actually be — throwing infrastructure at the problem gets you better eyes, not better judgment.
Honestly? For our use cases — document extraction, UI understanding, maybe some vision-heavy workflows — Muse Spark might quietly become a useful specialist. Just don't let it plan anything.
What To Do
Benchmark Muse Spark against GPT-4o and Gemini 1.5 Pro specifically on your document parsing or UI screenshot tasks — if it hits that 86% figure understanding on your data, the API pricing math might make it worth integrating as a specialist model.
Cited By
React