On April 8, Meta released Muse Spark.
It is Meta's first-ever fully closed-source, self-developed flagship AI model, and the first major result since the company spent $14 billion to bring in Alexandr Wang and establish the Meta Superintelligence Labs.
No Longer Llama
For years, Llama has been Meta's banner — open-source, locally deployable, and research-friendly.
Muse Spark takes a different path. It is only available on meta.ai and the Meta AI app, with a private API still in limited user testing. "We hope to open-source some version in the future" — the open-source community can probably hear the nuance in that statement.
The shift from Llama to Muse Spark is not accidental. By late 2025, Meta had already revealed two proprietary models, codenamed Avocado and Mango. Muse Spark is the first release ahead of those, and the direction is clear: Meta intends to compete head-on with OpenAI and Google on the closed-source track.
Most Notable Technology: Thought Compression
Muse Spark features a mode called Contemplating, roughly equivalent to OpenAI's Thinking mode — allowing the model to think deeply before responding.
But the interesting part lies in its internal training mechanism.
Meta engineers introduced a penalty during the reinforcement learning phase for the model's thinking time — thinking too much would cost points. Initially, the model responded by producing answers faster (with degraded quality). But as training continued, a turning point emerged: the model began to learn to compress its thinking — using fewer tokens while maintaining the same accuracy.
Meta calls this thought compression.
The result? According to Meta, Muse Spark requires more than an order of magnitude less computation than its predecessor flagship, Llama 4 Maverick, to achieve the same capability.
This is not a minor optimization; it is an order-of-magnitude difference.
Benchmark Performance
In Contemplating mode:
- Humanity's Last Exam: 58% (a difficult benchmark; 58% is a competitive score)
- FrontierScience Research: 38%
Compared to competitors of the same period: it can go toe-to-toe with Gemini Deep Think and GPT Pro, at least on these benchmarks.
Real-World Implications
Muse Spark will first power Meta AI features in Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban glasses. Meta's user base is in the billions, so this model will undergo real-world large-scale testing upon launch.
Of course, it is only available in the U.S. — this restriction means users in other regions cannot access it for now, but test data will emerge quickly.
With this move, Meta is essentially saying: we now have a good enough model that we no longer need to rely on an open-source strategy to attract a developer ecosystem.
Sources: Goodbye, Llama? Meta launches new proprietary AI model Muse Spark (VentureBeat); Meta debuts new AI model, attempting to catch Google, CocoLoop, OpenAI (CNBC); Introducing Muse Spark: Scaling Towards Personal Superintelligence (Meta AI Blog)