AI: Meta changes strategy with Muse Spark, the first model of its 'superintelligence' laboratory

Meta launched Muse Spark on Wednesday, April 8, the first model from Meta Superintelligence Labs, the unit led by Alexandr Wang. The model was developed in nine months after a complete overhaul of the company's AI infrastructure. It already powers the Meta AI assistant on the web and the dedicated app. It will be deployed on WhatsApp, Instagram, Facebook, and Messenger in the coming weeks.

Meta closes its AI after advocating for openness

The notable fact is not the model itself. It's the license. Muse Spark is proprietary, in stark contrast to the Llama series that Meta promoted for two years as the banner of 'open' AI. Except that this 'open' already deserved thick quotation marks. The Llama license has never been a strictly open-source license. It imposed commercial usage restrictions beyond 700 million monthly users and prohibited training of competing models. Far from the Apache 2.0 or MIT standards that the community considers truly free.

With Muse Spark, Meta even abandons this pretense of openness. The group 'hopes' to publish open-source versions in the future, without a schedule or commitment. The reason is clear: maintaining a proprietary advantage over OpenAI, Google, and Anthropic is worth more than the Llama developer ecosystem.

What Muse Spark can do, and what it still lacks

The model accepts text, voice, and images as input. It produces only text for now. Two modes coexist: a fast mode for simple queries and a 'Contemplating' mode that orchestrates several sub-agents in parallel. Meta claims superior efficiency compared to Llama 4 Maverick with ten times less computational power, thanks to a technique its researchers call 'reasoning compression'.

The benchmarks tell a nuanced story. On the Artificial Analysis Intelligence v4.0 index, Muse Spark ranks fourth with a score of 52. Ahead of it: Gemini 3.1 Pro and GPT-5.4 (57 each), then Claude Opus 4.6 (53). On the ARC AGI 2 abstract reasoning test, the gap widens: 42.5 versus 76.5 for Gemini. However, the model excels in visual graph comprehension: 86.4 on CharXiv. On the HealthBench Hard medical reasoning test, it reaches 42.8%, compared to 14.8% for Claude. For a team that has experienced hiring worthy of a transfer window followed by equally dramatic layoffs, this is a promising initial version that can be applauded.

Meta does not claim to have caught up with the competition on all fronts. The company is betting on specialization: health, shopping, visual understanding. Areas where the data from three billion users constitutes an advantage that neither OpenAI nor even Google can replicate. All financed by a massive investment envelope: between 115 and 135 billion dollars in infrastructure spending planned for 2026.

👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source: Meta