Meta Superintelligence Labs · 2026-04-08 · seismic
Muse Spark
First model from Meta Superintelligence Labs since the Alexandr Wang reshuffle — a natively multimodal reasoning model Meta claims matches Llama 4 Maverick with an order of magnitude less compute.
Meta's first model after the Alexandr Wang reshuffle — multimodal reasoning that matches Llama 4 Maverick with an order of magnitude less compute.
Key specs
| Humanity's last exam | 58% |
|---|---|
| Frontier science research | 38% |
| Compute vs llama 4 maverick | 10x less |
What is it?
Muse Spark is the first release from Meta Superintelligence Labs, the AI org Meta rebuilt around Alexandr Wang after the $14B Scale AI deal. It is a natively multimodal reasoning model with visual chain-of-thought and multi-agent orchestration, pitched as the model that will power consumer-facing AI inside Facebook, Instagram, WhatsApp, Messenger and the Ray-Ban Meta glasses. Unlike Llama, Muse Spark is closed-weights, accessed through meta.ai and an invite-only API.
How does it work?
Meta ships Muse Spark in two modes in meta.ai — Instant and Thinking — with a longer-reasoning Contemplating mode promised later. In Contemplating mode Meta reports 58% on Humanity's Last Exam and 38% on FrontierScience Research. The headline is efficiency: Meta says a rebuilt training stack lets Muse Spark match Llama 4 Maverick's capabilities with over an order of magnitude less training compute. Artificial Analysis independently scores it at 52, behind only Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6.
Why does it matter?
This is the first concrete proof that Meta's expensive bet on Wang and MSL is producing real models, not slides. The story is not the raw benchmarks — it is the efficiency. If MSL can keep delivering frontier-comparable models at 10x lower compute, Meta can ship them on phones and AR glasses at a cost no competitor can match.
Who is it for?
Anyone watching the consumer-AI race or building edge / on-device experiences.
Try it
meta.ai (US only at launch)